[ { "number": 1, "id": "baby/vistarc_q3_mos3_branch123_width320_ffn1120_heads10_20261009", "original_name": "320・MoS3・枝別1/2/3層混合", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": false, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.715, "comparison_scope": "Matched ages and observed peaks separately, including ordinary320 shared-last2 and all selected historical models" }, "milestones_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 44, "sources": [ "s3ac2028a03295e", "sc9ced036cdce83", "s557e5d189f5732", "sc1082b3d74ec9b", "sc3531f0ecf93b5", "s8ad80332a1a75c", "s951debf6f53be8", "s2f1550e4927224", "s90d46b431241aa", "s44f366cb4aea62", "s35f135453f1b5d" ], "points": [ { "age": 40.0, "six": 48.23083333333334, "five": 50.45, "tasks": { "blimp_filtered": 69.39, "supplement_filtered": 58.92, "ewok_filtered": 51.27, "entity_tracking": 19.98, "comps": 52.69, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 50.0 }, "checkpoint": "73028bf90313312eb6d926811f4d626e6b8693a5fdddcb3d29b2ae5b45e29d98", "precision": "fp32", "verification": "primary", "sources": [ "s2261a582a56657", "s358f18f1204f68" ] }, { "age": 50, "six": 48.73916666666667, "five": 51.654, "tasks": { "blimp_filtered": 69.31, "supplement_filtered": 62.95, "ewok_filtered": 51.91, "entity_tracking": 21.75, "comps": 52.35, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "a44859309aea9f522857251aa2a726234b8c98882c28a6f537be3b363b84d1eb", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc49cc717d6ef29", "s7a95b17f04d915", "s1a0a36b43ca3ed", "sa77bc22deb09d2", "s647f12ed14d347", "se5ace5456c8dd7", "sc432a1c31c0c38", "s59a3c5f0f4742d" ] }, { "age": 60, "six": 48.51416666666667, "five": 51.378, "tasks": { "blimp_filtered": 69.2, "supplement_filtered": 61.37, "ewok_filtered": 51.47, "entity_tracking": 22.3, "comps": 52.55, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 48.0 }, "checkpoint": "b7f648426d23083406a80bbf3f1ca889f9c63029548bc9b31776f041e84f2abe", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s55b2d319d5ba8b", "s45b1d93c52c823", "sfc0df3c4871be4", "sa77bc22deb09d2", "s647f12ed14d347", "se5ace5456c8dd7", "sc432a1c31c0c38", "s59a3c5f0f4742d" ] }, { "age": 70, "six": 48.82083333333333, "five": 51.352, "tasks": { "blimp_filtered": 69.83, "supplement_filtered": 60.9, "ewok_filtered": 51.04, "entity_tracking": 22.37, "comps": 52.62, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 50.0 }, "checkpoint": "ed25fa11197f1368a3982e240f62be17e85e80d0deb32e6ebcad5f9091ad5f00", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5b58e68456be20", "s85aca586603e64", "sb343124a51eb5d", "sa77bc22deb09d2", "s647f12ed14d347", "se5ace5456c8dd7", "sc432a1c31c0c38", "s59a3c5f0f4742d" ] }, { "age": 80, "six": 49.8675, "five": 52.002, "tasks": { "blimp_filtered": 69.45, "supplement_filtered": 63.84, "ewok_filtered": 51.01, "entity_tracking": 23.33, "comps": 52.38, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 58.0 }, "checkpoint": "3250acb4a4eba28881754f43db5be762745738adddc94c1d5480794d31d3ce38", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s8929ce6365c999", "sf2b6d20d9796d5", "se6bb683ff4d433", "sa77bc22deb09d2", "s647f12ed14d347", "se5ace5456c8dd7", "sc432a1c31c0c38", "s59a3c5f0f4742d" ] }, { "age": 90, "six": 48.964166666666664, "five": 51.924, "tasks": { "blimp_filtered": 70.43, "supplement_filtered": 61.52, "ewok_filtered": 52.36, "entity_tracking": 22.78, "comps": 52.53, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "7dd39caa9337951837c2d87b336833dc82caac7f19a72f0176a92b1efeee3a9c", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s76e25bb3565cf4", "sc39b873a064777", "s393e4545fddf69", "sa77bc22deb09d2", "s647f12ed14d347", "se5ace5456c8dd7", "sc432a1c31c0c38", "s59a3c5f0f4742d" ] }, { "age": 100, "six": 49.0625, "five": 51.354, "tasks": { "blimp_filtered": 69.64, "supplement_filtered": 60.71, "ewok_filtered": 51.77, "entity_tracking": 21.89, "comps": 52.76, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 49.0 }, "checkpoint": "2283ee6cf9c4062c3ffd3012104968329c910fb77541312b3fc7354e2ab0ac25", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s8a2e03ea2a6017", "sa7e5b1c6719010", "s6e97812287e78f", "sa77bc22deb09d2", "s647f12ed14d347", "se5ace5456c8dd7", "sc432a1c31c0c38", "s59a3c5f0f4742d" ] } ], "peak": { "five": { "score": 52.002, "ages": [ 80 ], "rank": 1 }, "six": { "score": 49.8675, "ages": [ 80 ], "rank": 1 } }, "clear_points": [ { "number": 1, "age": 80, "seed": 20260907, "checkpoint_sha256": "3250acb4a4eba28881754f43db5be762745738adddc94c1d5480794d31d3ce38", "old_six": 49.8675, "old_five": 52.002, "old_tasks": { "blimp_filtered": 69.45, "supplement_filtered": 63.84, "ewok_filtered": 51.01, "entity_tracking": 23.33, "comps": 52.38, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 58.0 }, "correct": 12269, "total": 17274, "accuracy": 71.02581915016788, "ties": 0, "tasks": { "blimp_filtered": { "correct": 7001, "total": 8392, "ties": 0 }, "comps": { "correct": 3822, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 0, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 15, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1026, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 72, "total": 114, "ties": 0 }, "boolq": { "correct": 8, "total": 53, "ties": 0 }, "mnli": { "correct": 48, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 63.38773906999995, "checkpoint_rank": 7, "rank": 7 } ], "clear_peak": { "number": 1, "age": 80, "seed": 20260907, "checkpoint_sha256": "3250acb4a4eba28881754f43db5be762745738adddc94c1d5480794d31d3ce38", "old_six": 49.8675, "old_five": 52.002, "old_tasks": { "blimp_filtered": 69.45, "supplement_filtered": 63.84, "ewok_filtered": 51.01, "entity_tracking": 23.33, "comps": 52.38, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 58.0 }, "correct": 12269, "total": 17274, "accuracy": 71.02581915016788, "ties": 0, "tasks": { "blimp_filtered": { "correct": 7001, "total": 8392, "ties": 0 }, "comps": { "correct": 3822, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 0, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 15, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1026, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 72, "total": 114, "ties": 0 }, "boolq": { "correct": 8, "total": 53, "ties": 0 }, "mnli": { "correct": 48, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 63.38773906999995, "checkpoint_rank": 7, "rank": 7 }, "present": true }, { "number": 2, "id": "baby/vistarc_q3_mos3_last2_width384_ffn1344_heads12_20261008", "original_name": "384・MoS3・2層混合", "title": "384幅・FFN1344・9層", "topology": "MoS3・共通2層混合", "recipe": "主要0.0005333／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1344, "depth": 9, "parts": { "width": 384, "ffn": 1344, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.0005333333333333334, 0.00046666666666666666, 0.00039999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1344, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last2", "attention_received_gate": true, "output_components": 3, "output_state_layers": 2, "reference_attention_width": 256, "ffn_additive_bypass": false, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3_LAST2 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 65536, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.315, "comparison_scope": "last2_256 parent plus interrupted last2_272 and last3/last4 references" }, "milestones_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 52, "sources": [ "sf2d569f7bfa96a", "sd3fd1bb595cec1", "sde0b412fea7ecd", "s8cb286657386e7", "sdfa156d2f8b156", "s0dade360afdfc3" ], "points": [ { "age": 60, "six": 48.54416666666667, "five": 50.32, "tasks": { "blimp_filtered": 68.52, "supplement_filtered": 58.29, "ewok_filtered": 50.97, "entity_tracking": 21.49, "comps": 52.33, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 57.0 }, "checkpoint": "0f989084c1f60e061355301438da23db6cf43ac84c8aa7628b0251819632c5a2", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s33e8c0adbc195e", "s4a90425a15e07f", "s8a77b8572e986f", "saf4339eafc7aa7", "sea2d7867b2a076", "sbfed5e1e31342c" ] }, { "age": 70, "six": 48.983333333333334, "five": 51.138, "tasks": { "blimp_filtered": 69.82, "supplement_filtered": 59.76, "ewok_filtered": 52.1, "entity_tracking": 21.2, "comps": 52.81, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 57.0 }, "checkpoint": "665b10314f3ca03df6b8c64d778cf029f8c99638ae52db0f26125130a386b8e2", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s8965d868e4c21c", "sb3f25cb6279a18", "s8a77b8572e986f", "saf4339eafc7aa7", "sea2d7867b2a076", "sbfed5e1e31342c" ] }, { "age": 80, "six": 49.715, "five": 51.04, "tasks": { "blimp_filtered": 68.98, "supplement_filtered": 61.22, "ewok_filtered": 51.03, "entity_tracking": 21.62, "comps": 52.35, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 59.0 }, "checkpoint": "c8e66041d2f8c7e5b71346e69b69790d8355e4883530bc6c82418dc8d9ef0862", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s824ae5b3c41da9", "s6cfbb3d92d9432", "s8a77b8572e986f", "saf4339eafc7aa7", "sea2d7867b2a076", "sbfed5e1e31342c" ] }, { "age": 90, "six": 48.094166666666666, "five": 50.186, "tasks": { "blimp_filtered": 68.1, "supplement_filtered": 58.26, "ewok_filtered": 52.38, "entity_tracking": 19.85, "comps": 52.34, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 51.0 }, "checkpoint": "ff850a611adfe31dbe857ab4f2a59124327a12f21be74dd8f1736c4ff9ca99e9", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s96cfea32beef15", "s3507c8545582bf", "s8a77b8572e986f", "saf4339eafc7aa7", "sea2d7867b2a076", "sbfed5e1e31342c" ] }, { "age": 100, "six": 48.428333333333335, "five": 51.078, "tasks": { "blimp_filtered": 68.69, "supplement_filtered": 59.63, "ewok_filtered": 53.69, "entity_tracking": 20.68, "comps": 52.7, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 49.0 }, "checkpoint": "1f943119223e57c9ef8046fd884ef12670937926c7c4ddf4fd754419f9efd296", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "se9ff92ea5500bd", "s162be4df1e1ab0", "s8a77b8572e986f", "saf4339eafc7aa7", "sea2d7867b2a076", "sbfed5e1e31342c" ] } ], "peak": { "five": { "score": 51.138, "ages": [ 70 ], "rank": 30 }, "six": { "score": 49.715, "ages": [ 80 ], "rank": 3 } }, "clear_points": [ { "number": 2, "age": 80, "seed": 20260907, "checkpoint_sha256": "c8e66041d2f8c7e5b71346e69b69790d8355e4883530bc6c82418dc8d9ef0862", "old_six": 49.715, "old_five": 51.04, "old_tasks": { "blimp_filtered": 68.98, "supplement_filtered": 61.22, "ewok_filtered": 51.03, "entity_tracking": 21.62, "comps": 52.35, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 59.0 }, "correct": 12185, "total": 17274, "accuracy": 70.53953919184902, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6893, "total": 8392, "ties": 0 }, "comps": { "correct": 3887, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 16, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1002, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 2, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 1, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 3, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 75, "total": 114, "ties": 0 }, "boolq": { "correct": 10, "total": 53, "ties": 0 }, "mnli": { "correct": 25, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 213, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 65.94253349799692, "checkpoint_rank": 16, "rank": 14 }, { "number": 2, "age": 70, "seed": 20260907, "checkpoint_sha256": "665b10314f3ca03df6b8c64d778cf029f8c99638ae52db0f26125130a386b8e2", "old_six": 48.983333333333334, "old_five": 51.138, "old_tasks": { "blimp_filtered": 69.82, "supplement_filtered": 59.76, "ewok_filtered": 52.1, "entity_tracking": 21.2, "comps": 52.81, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 57.0 }, "correct": 12104, "total": 17274, "accuracy": 70.0706263748987, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6819, "total": 8392, "ties": 0 }, "comps": { "correct": 3902, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 17, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 982, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 70, "total": 114, "ties": 0 }, "boolq": { "correct": 11, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 66.22164592400077, "checkpoint_rank": 19 } ], "clear_peak": { "number": 2, "age": 80, "seed": 20260907, "checkpoint_sha256": "c8e66041d2f8c7e5b71346e69b69790d8355e4883530bc6c82418dc8d9ef0862", "old_six": 49.715, "old_five": 51.04, "old_tasks": { "blimp_filtered": 68.98, "supplement_filtered": 61.22, "ewok_filtered": 51.03, "entity_tracking": 21.62, "comps": 52.35, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 59.0 }, "correct": 12185, "total": 17274, "accuracy": 70.53953919184902, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6893, "total": 8392, "ties": 0 }, "comps": { "correct": 3887, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 16, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1002, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 2, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 1, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 3, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 75, "total": 114, "ties": 0 }, "boolq": { "correct": 10, "total": 53, "ties": 0 }, "mnli": { "correct": 25, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 213, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 65.94253349799692, "checkpoint_rank": 16, "rank": 14 }, "present": true }, { "number": 3, "id": "run/IPRLM384_LR_B_MAJOR90_OTHER105_SEED20260907_40M_20261020", "original_name": "IPRLM384_LR_B_MAJOR90_OTHER105_SEED20260907_40M_20261020", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要−10％・その他＋5％", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All15matchedseed40MLRconditions;reuse12controls;allgenres/pairedoutcomes/fullinventory" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 119, "sources": [ "s4c3b8439c9b1f9" ], "points": [ { "age": 40.0, "six": 48.079166666666666, "five": 50.474, "tasks": { "blimp_filtered": 68.98, "supplement_filtered": 58.84, "ewok_filtered": 52.23, "entity_tracking": 19.61, "comps": 52.71, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 46.0 }, "checkpoint": "312ed369d95cd5a414916c445b54f9bd575c2ce3ecd2626af6537498a897ea20", "precision": "fp32", "verification": "primary", "sources": [ "s85dc68bb33dba2", "s03150237e4bcaf" ] }, { "age": 70.0, "six": 49.641666666666666, "five": 51.354, "tasks": { "blimp_filtered": 70.15, "supplement_filtered": 59.76, "ewok_filtered": 53.45, "entity_tracking": 21.07, "comps": 52.34, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 54.0 }, "checkpoint": "ccda1fea95a2eb157d8bb0c3ed020ea16127a48c087e907e592310ad66d6c864", "precision": "fp32", "verification": "primary", "sources": [ "s36050253956dd0", "s03150237e4bcaf" ] }, { "age": 80.0, "six": 48.71333333333333, "five": 50.732, "tasks": { "blimp_filtered": 69.36, "supplement_filtered": 60.61, "ewok_filtered": 52.4, "entity_tracking": 19.22, "comps": 52.07, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 52.0 }, "checkpoint": "14320930026a87ce0b09b569446e821098e6a18d75d4b1e0a50007a1bddb8562", "precision": "fp32", "verification": "primary", "sources": [ "s915dfe21488743", "s03150237e4bcaf" ] } ], "peak": { "five": { "score": 51.354, "ages": [ 70.0 ], "rank": 16 }, "six": { "score": 49.641666666666666, "ages": [ 70.0 ], "rank": 4 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 4, "id": "run/IPRLM_FFN1121_320_SEED731946281_100M_20261014", "original_name": "IPRLM_FFN1121_320_SEED731946281_100M_20261014", "title": "320幅・FFN1121・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 320, "ffn": 1121, "depth": 9, "parts": { "width": 320, "ffn": 1121, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.503125 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1121, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_preserve_parent_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes" ], "intermediate": "After100M train:70/80/90/100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "selection_metric": "five_axis", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "no_endpoint_penalty": true }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 153, "sources": [ "s039fda49ce45a7", "s1ec03a13341e65", "s17d54a96dd90e1" ], "points": [ { "age": 70.0, "six": 47.8675, "five": 50.402, "tasks": { "blimp_filtered": 68.62, "supplement_filtered": 60.28, "ewok_filtered": 51.0, "entity_tracking": 19.87, "comps": 52.24, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 50.0 }, "checkpoint": "6902ff774a2a665c886d0bdf5d4bdbe818846cec1c8161bf795074558eb2ecac", "precision": "fp32", "verification": "primary", "sources": [ "s5b6538f28e7292", "s33d9602fe7fd1f" ] }, { "age": 80.0, "six": 49.465, "five": 50.542, "tasks": { "blimp_filtered": 68.33, "supplement_filtered": 60.63, "ewok_filtered": 51.63, "entity_tracking": 19.49, "comps": 52.63, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 60.0 }, "checkpoint": "238318e11067268ac8e7a691283dd2de8632fb8268538cfa645157b6426fed9a", "precision": "fp32", "verification": "primary", "sources": [ "sca83be2d6f054c", "s01b0527570607b" ] }, { "age": 90.0, "six": 48.410833333333336, "five": 50.466, "tasks": { "blimp_filtered": 68.52, "supplement_filtered": 59.0, "ewok_filtered": 52.06, "entity_tracking": 19.95, "comps": 52.8, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 52.0 }, "checkpoint": "c0d6b2fe6a73a8499f14a53d806720c04bac98c5e4ce048bea783377f504bae8", "precision": "fp32", "verification": "primary", "sources": [ "sd3f7246db0a34e", "sbdc9638bd3be70" ] }, { "age": 100.0, "six": 48.428333333333335, "five": 50.29, "tasks": { "blimp_filtered": 67.98, "supplement_filtered": 58.85, "ewok_filtered": 52.29, "entity_tracking": 19.64, "comps": 52.69, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 53.0 }, "checkpoint": "96a6de6adad62d6b2879af7eae7b61027277d2f24c471c3e1934489dd68eb1a4", "precision": "fp32", "verification": "primary", "sources": [ "sa012bd7f8670c2", "s9c0f581ca494d5" ] } ], "peak": { "five": { "score": 50.542, "ages": [ 80.0 ], "rank": 106 }, "six": { "score": 49.465, "ages": [ 80.0 ], "rank": 6 } }, "clear_points": [ { "number": 4, "age": 80, "seed": 731946281, "checkpoint_sha256": "238318e11067268ac8e7a691283dd2de8632fb8268538cfa645157b6426fed9a", "old_six": 49.465, "old_five": 50.542, "old_tasks": { "blimp_filtered": 68.33, "supplement_filtered": 60.63, "ewok_filtered": 51.63, "entity_tracking": 19.49, "comps": 52.63, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 60.0 }, "correct": 12196, "total": 17274, "accuracy": 70.6032187102003, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6912, "total": 8392, "ties": 0 }, "comps": { "correct": 3904, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 1, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 12, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 986, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 1, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 1, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 0, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 0, "total": 2, "ties": 0 }, "piqa1838": { "correct": 74, "total": 114, "ties": 0 }, "boolq": { "correct": 9, "total": 53, "ties": 0 }, "mnli": { "correct": 28, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 62.14728917100001, "checkpoint_rank": 13, "rank": 12 } ], "clear_peak": { "number": 4, "age": 80, "seed": 731946281, "checkpoint_sha256": "238318e11067268ac8e7a691283dd2de8632fb8268538cfa645157b6426fed9a", "old_six": 49.465, "old_five": 50.542, "old_tasks": { "blimp_filtered": 68.33, "supplement_filtered": 60.63, "ewok_filtered": 51.63, "entity_tracking": 19.49, "comps": 52.63, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 60.0 }, "correct": 12196, "total": 17274, "accuracy": 70.6032187102003, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6912, "total": 8392, "ties": 0 }, "comps": { "correct": 3904, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 1, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 12, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 986, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 1, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 1, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 0, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 0, "total": 2, "ties": 0 }, "piqa1838": { "correct": 74, "total": 114, "ties": 0 }, "boolq": { "correct": 9, "total": 53, "ties": 0 }, "mnli": { "correct": 28, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 62.14728917100001, "checkpoint_rank": 13, "rank": 12 }, "present": true }, { "number": 5, "id": "run/Q3_MOS3_BRANCH123_10L_WIDTH320_FFN1120_HEADS10_20261011", "original_name": "Q3_MOS3_BRANCH123_10L_WIDTH320_FFN1120_HEADS10_20261011", "title": "320幅・FFN1120・10層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 10, "parts": { "width": 320, "ffn": 1120, "depth": 10, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "10層 / 9・10層 / 8・9・10層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 10 ], [ 9, 10 ], [ 8, 9, 10 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 10, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_init_only", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 10 ], [ 9, 10 ], [ 8, 9, 10 ] ], "initialization_layers": 9, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes70/80/90/100M", "SyntaxGym31+PIQA1838 at70/80/90/100M", "frozen bAbI1947+ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 TF32off exact prefix graph; actual10-layer320 speed, score parity and all capacity buckets verified", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Matched ages and observed peaks separately, including ordinary320 shared-last2 and all selected historical models" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ], "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 144, "sources": [ "s6c540e2e6d34ac", "sbd9ddbbd39bf6e", "s4cea40920f296b", "s9dfbb00c20aaba", "s21fde1fd50e1b1", "s86495935b598f8" ], "points": [ { "age": 40.0, "six": 47.76833333333333, "five": 49.792, "tasks": { "blimp_filtered": 67.95, "supplement_filtered": 57.82, "ewok_filtered": 51.88, "entity_tracking": 19.12, "comps": 52.19, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 52.0 }, "checkpoint": "8d4cecd8ba64c042201d4a2f50d23df1f506ac80421f3590ce05bd4f5d23e58d", "precision": "fp32", "verification": "primary", "sources": [ "s84a0ca6f4dbff7", "se05ac28beb4e0e" ] }, { "age": 70.0, "six": 48.575, "five": 50.56, "tasks": { "blimp_filtered": 69.9, "supplement_filtered": 59.82, "ewok_filtered": 51.29, "entity_tracking": 19.55, "comps": 52.24, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 54.0 }, "checkpoint": "4b42b5028b8cafec125630f3b32c67fd6651ea655509d5ca9535c55a041735b4", "precision": "fp32", "verification": "primary", "sources": [ "sb8d06e8e0af454", "sf199a973b1bcfb", "s0fc4af8aa296ff", "s688da1c8dbf835", "s629b3ca14022b4", "s01294abc9a664b" ] }, { "age": 80.0, "six": 49.44833333333333, "five": 51.008, "tasks": { "blimp_filtered": 69.05, "supplement_filtered": 60.99, "ewok_filtered": 51.97, "entity_tracking": 20.75, "comps": 52.28, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 60.0 }, "checkpoint": "eef0c1841ef65c18530518c1f698753cdf4c6469add466deaea816638c9839b4", "precision": "fp32", "verification": "primary", "sources": [ "s055f85838c682a", "se36118db55faa6", "s0fc4af8aa296ff", "s688da1c8dbf835", "s629b3ca14022b4", "s01294abc9a664b" ] }, { "age": 90.0, "six": 48.14333333333333, "five": 49.854, "tasks": { "blimp_filtered": 68.49, "supplement_filtered": 58.26, "ewok_filtered": 51.35, "entity_tracking": 19.21, "comps": 51.96, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 52.0 }, "checkpoint": "a85afdbf4b7470d1f11c66e55a463eeb8ca80466447f97e617ce7a3be669be8b", "precision": "fp32", "verification": "primary", "sources": [ "s6c32b7e1280b89", "s0368c69f24f196", "s0fc4af8aa296ff", "s688da1c8dbf835", "s629b3ca14022b4", "s01294abc9a664b" ] }, { "age": 100.0, "six": 47.945, "five": 49.904, "tasks": { "blimp_filtered": 67.92, "supplement_filtered": 58.96, "ewok_filtered": 51.05, "entity_tracking": 19.32, "comps": 52.27, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 53.0 }, "checkpoint": "e191e98a6a0500fcc0344d95f4a79043c7e08a269cf3ecf82533021b68ff0ebc", "precision": "fp32", "verification": "primary", "sources": [ "s7a1467d561ea2d", "sc34a5e809f3035", "s0fc4af8aa296ff", "s688da1c8dbf835", "s629b3ca14022b4", "s01294abc9a664b" ] } ], "peak": { "five": { "score": 51.008, "ages": [ 80.0 ], "rank": 45 }, "six": { "score": 49.44833333333333, "ages": [ 80.0 ], "rank": 7 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 6, "id": "run/IPRLM384_JOINT_NORM820_PAIR_SEED20260907_40M_20261021", "original_name": "IPRLM384_JOINT_NORM820_PAIR_SEED20260907_40M_20261021", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "Attention前・FFN前Norm820", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All20same-seed40Mconditions;all7tasks,5/6aggregates,pairedoutcomes;fullinventory;deltaagainstfailedjointparent" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 110, "sources": [ "s546e09109ae8b4", "se6e76ff14ed56e" ], "points": [ { "age": 40.0, "six": 48.21083333333333, "five": 50.714, "tasks": { "blimp_filtered": 69.28, "supplement_filtered": 59.34, "ewok_filtered": 52.79, "entity_tracking": 19.8, "comps": 52.36, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 51.0 }, "checkpoint": "16971237c9ce61ecb12fbdf28d8cd707a3687b4a5c1604c0b11cf816fc3e0577", "precision": "fp32", "verification": "primary", "sources": [ "se84b34fb276482", "s95e4ab81f4f78f" ] }, { "age": 70.0, "six": 49.314166666666665, "five": 51.044, "tasks": { "blimp_filtered": 69.29, "supplement_filtered": 59.7, "ewok_filtered": 52.19, "entity_tracking": 21.32, "comps": 52.72, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 59.0 }, "checkpoint": "d94138f1474c0276c1e19468da029bf01e83027776e7ec983837724da68bab35", "precision": "fp32", "verification": "primary", "sources": [ "s70442b1fd4e898", "s95e4ab81f4f78f" ] }, { "age": 80.0, "six": 49.41583333333333, "five": 51.366, "tasks": { "blimp_filtered": 69.44, "supplement_filtered": 61.16, "ewok_filtered": 52.41, "entity_tracking": 21.14, "comps": 52.68, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 57.0 }, "checkpoint": "aba6fd48d302eb71c34a1b7e2fe79157b247bfd7477516965f1fa73ec1d1598f", "precision": "fp32", "verification": "primary", "sources": [ "sf2e7704ae1b0a0", "s95e4ab81f4f78f" ] } ], "peak": { "five": { "score": 51.366, "ages": [ 80.0 ], "rank": 14 }, "six": { "score": 49.41583333333333, "ages": [ 80.0 ], "rank": 8 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 7, "id": "baby/vistarc_q3_mos3_layerwise987_width320_ffn1120_heads10_20261009", "original_name": "320・MoS3・各層分離9/8/7", "title": "320幅・FFN1120・9層", "topology": "MoS3・各層分離", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8層 / 7層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "各層を個別", "reference_layers": [ [ 9 ], [ 8 ], [ 7 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_layerwise987", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8 ], [ 7 ] ], "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph; exact single-layer branch outputs. GPU parity, speed and cache capacity checked after predecessor completes.", "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Separated320 versus its nested320 parent at matched50/60/70/80/90/100M and observed peaks; full historical references retained." }, "milestones_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 55, "sources": [ "scebaabf6c812b1", "s15fd028dae40f1", "s0246e6b7029f31", "s8ee9b53f18e47a", "s1736422d6ae2c6", "s4b9b0cf07166d4", "s8896b3ce405c14", "s5e080d249a883c", "s6306429110761e" ], "points": [ { "age": 40.0, "six": 48.27166666666667, "five": 50.796, "tasks": { "blimp_filtered": 69.97, "supplement_filtered": 58.14, "ewok_filtered": 52.59, "entity_tracking": 20.49, "comps": 52.79, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 48.0 }, "checkpoint": "b5c59338667d34888f1ac78f35d40beb9905a8757ed8b1432fe0918d1fd32a95", "precision": "fp32", "verification": "primary", "sources": [ "sbf7de57e63e466", "sab193196954c65" ] }, { "age": 50, "six": 47.99166666666667, "five": 50.954, "tasks": { "blimp_filtered": 70.1, "supplement_filtered": 58.56, "ewok_filtered": 52.61, "entity_tracking": 20.9, "comps": 52.6, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 45.0 }, "checkpoint": "8e30742189f78f4f3c07a9d3b4d3073fe0323591d6f54bff67526faf4ea4da4f", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s3e021344569e8c", "s543f2146133f0f", "s542eab41f45a63", "s0ba6428c5552d7", "s2b68d821d1cde0", "sfe01bfdab8cfeb" ] }, { "age": 60, "six": 47.72, "five": 49.934, "tasks": { "blimp_filtered": 67.14, "supplement_filtered": 57.83, "ewok_filtered": 50.81, "entity_tracking": 21.99, "comps": 51.9, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "3f27fcc18ef762e236b194ab09e55cf143c2ebf52baaefcfdfb456381b9ca6af", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sf608bd0d3c76fa", "s5cfe74e65fd877", "s542eab41f45a63", "s0ba6428c5552d7", "s2b68d821d1cde0", "sfe01bfdab8cfeb" ] }, { "age": 70, "six": 48.04333333333334, "five": 51.004, "tasks": { "blimp_filtered": 69.12, "supplement_filtered": 59.13, "ewok_filtered": 52.88, "entity_tracking": 21.8, "comps": 52.09, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 49.0 }, "checkpoint": "a12192029602f07c987675ba73188ffbff44d3997080eb069f8231d3937edf91", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4809c133efb805", "s4fa5845e32d4fa", "s542eab41f45a63", "s0ba6428c5552d7", "s2b68d821d1cde0", "sfe01bfdab8cfeb" ] }, { "age": 80, "six": 49.401666666666664, "five": 51.646, "tasks": { "blimp_filtered": 69.08, "supplement_filtered": 60.41, "ewok_filtered": 53.87, "entity_tracking": 22.44, "comps": 52.43, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 55.0 }, "checkpoint": "12a58605e40644c8cd60280c5a9652758268a7ed72aa1cad7eea075c5eb3035f", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s28c8a28f84f1fb", "s29d744d49541a3", "s542eab41f45a63", "s0ba6428c5552d7", "s2b68d821d1cde0", "sfe01bfdab8cfeb" ] }, { "age": 90, "six": 48.45333333333333, "five": 50.714, "tasks": { "blimp_filtered": 68.29, "supplement_filtered": 57.36, "ewok_filtered": 54.29, "entity_tracking": 21.39, "comps": 52.24, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "68d8700a399e5ff471d34df67b088f725c65e4079a105118811b64a19775a232", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "se753dee3afc3bc", "sabe66586eeeeea", "s542eab41f45a63", "s0ba6428c5552d7", "s2b68d821d1cde0", "sfe01bfdab8cfeb" ] }, { "age": 100, "six": 48.66166666666667, "five": 51.158, "tasks": { "blimp_filtered": 68.29, "supplement_filtered": 59.43, "ewok_filtered": 54.14, "entity_tracking": 21.64, "comps": 52.29, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 51.0 }, "checkpoint": "be54f14889a7f2198857adcd6560b38af8e7a2cbe20327d684e951ea24294c26", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4a342f4d95954b", "sf1e479fe40f2ff", "s542eab41f45a63", "s0ba6428c5552d7", "s2b68d821d1cde0", "sfe01bfdab8cfeb" ] } ], "peak": { "five": { "score": 51.646, "ages": [ 80 ], "rank": 4 }, "six": { "score": 49.401666666666664, "ages": [ 80 ], "rank": 9 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 8, "id": "baby/vistarc_q3_mos3_last1_width384_ffn1344_heads12_20261008", "original_name": "384・MoS3・1層参照", "title": "384幅・FFN1344・9層", "topology": "MoS3", "recipe": "主要0.0005333／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1344, "depth": 9, "parts": { "width": 384, "ffn": 1344, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "最終層", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.0005333333333333334, 0.00046666666666666666, 0.00039999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1344, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last1", "attention_received_gate": true, "output_components": 3, "output_state_layers": 1, "reference_attention_width": 256, "ffn_additive_bypass": false, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph; verify last1 against strict recurrence", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 65536, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.138, "same_family_peak_six_axis": 49.315, "comparison_scope": "last2_384 parent and prior256/240/272 reference models", "parent_peak_six_axis": 49.715 }, "milestones_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 47, "sources": [ "sbc34822d55323c", "s499816fb23735c", "sbab8f63b9d99df", "s474d583e9029b6", "sbcff04df876958", "s73cc7aa54d3076" ], "points": [ { "age": 60, "six": 47.54416666666667, "five": 50.326, "tasks": { "blimp_filtered": 69.15, "supplement_filtered": 58.07, "ewok_filtered": 50.59, "entity_tracking": 21.47, "comps": 52.35, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 43.0 }, "checkpoint": "622837c25c0e35281fd8ec77071e18736ad84c50e8c6c47462461d565512e7ff", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s07821b35cf3999", "s51298d76fe8245", "s49f47f1fd74afc", "s4ca547afe71afc", "s09f1150e04fbb3", "s5da6d8f207ec86" ] }, { "age": 70, "six": 49.37916666666667, "five": 51.328, "tasks": { "blimp_filtered": 70.49, "supplement_filtered": 58.45, "ewok_filtered": 54.87, "entity_tracking": 20.11, "comps": 52.72, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 55.0 }, "checkpoint": "f99d1eab9e32feab66df5ca06888dbfd0f6fb2f23eb3240a581856e9e68db811", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sa9be90498897c2", "se2e6d0348f8ae8", "s49f47f1fd74afc", "s4ca547afe71afc", "s09f1150e04fbb3", "s5da6d8f207ec86" ] }, { "age": 80, "six": 48.901666666666664, "five": 51.046, "tasks": { "blimp_filtered": 69.46, "supplement_filtered": 59.89, "ewok_filtered": 52.5, "entity_tracking": 20.45, "comps": 52.93, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 55.0 }, "checkpoint": "92e3c9ec3463b6b3331360397488ed9fa03abd1adcd5b2685f45fed26875c371", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sb80cd3cd4c99f3", "s6e64ba30c57f5f", "s49f47f1fd74afc", "s4ca547afe71afc", "s09f1150e04fbb3", "s5da6d8f207ec86" ] }, { "age": 90, "six": 48.815, "five": 50.848, "tasks": { "blimp_filtered": 69.26, "supplement_filtered": 60.0, "ewok_filtered": 52.24, "entity_tracking": 20.46, "comps": 52.28, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 54.0 }, "checkpoint": "741326a9635a0fb2378896e76f38dede947196c1a88306ba0a92bde5e1b6fb59", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s266b9b9328c8fe", "sf23280c23b3406", "s49f47f1fd74afc", "s4ca547afe71afc", "s09f1150e04fbb3", "s5da6d8f207ec86" ] }, { "age": 100, "six": 48.535833333333336, "five": 51.116, "tasks": { "blimp_filtered": 69.91, "supplement_filtered": 60.54, "ewok_filtered": 51.51, "entity_tracking": 21.07, "comps": 52.55, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 47.0 }, "checkpoint": "0b5a18daf1971a0705de602695ab16943297b060fe4c930d8b62b7ab83c749fc", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sef614610731e6d", "s1475530e91677f", "s49f47f1fd74afc", "s4ca547afe71afc", "s09f1150e04fbb3", "s5da6d8f207ec86" ] } ], "peak": { "five": { "score": 51.328, "ages": [ 70 ], "rank": 18 }, "six": { "score": 49.37916666666667, "ages": [ 70 ], "rank": 10 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 9, "id": "run/IPRLM320_FFN1121_PRESERVE_PARENT_INIT_20261013", "original_name": "IPRLM320_FFN1121_PRESERVE_PARENT_INIT_20261013", "title": "320幅・FFN1121・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1121, "depth": 9, "parts": { "width": 320, "ffn": 1121, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.503125 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1121, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_preserve_parent_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at70/80/90/100M", "SyntaxGym31 and originalPIQA1838 at70/80/90/100M", "frozen bAbI1947 and ProofWriter1000 at70/80/90/100M" ], "intermediate": "No evaluation during training; all benchmarks after100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Original parent320 and child matched70/80/90/100M, with observed peaks separately" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 74, "sources": [ "sb842c7898e6f6a", "sa34c1e11a19a6d", "s578734bcccea5f", "s914883e2483d72", "sc5f3fbf594090b", "s1ae0b4496f4d3d", "s0e8b79c89c9c9d", "s2f28e70ff4118a", "s86d3baf567078e" ], "points": [ { "age": 40.0, "six": 48.27166666666667, "five": 50.408, "tasks": { "blimp_filtered": 69.89, "supplement_filtered": 57.46, "ewok_filtered": 51.61, "entity_tracking": 20.91, "comps": 52.17, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 48.0 }, "checkpoint": "78cbcf0ce108d5750a43b23b99f7d2aa2220e84679934c2a348681e297d08050", "precision": "fp32", "verification": "primary", "sources": [ "s26394aae73693d", "s2b5c519de5b741" ] }, { "age": 70.0, "six": 48.4225, "five": 50.78, "tasks": { "blimp_filtered": 69.7, "supplement_filtered": 59.68, "ewok_filtered": 51.14, "entity_tracking": 20.88, "comps": 52.5, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 49.0 }, "checkpoint": "6506d8987da1a125702205f351928d13ef2ed178762cb831c82d6ab4434a75de", "precision": "fp32", "verification": "primary", "sources": [ "s0ac8ac5ee72a8f", "s24e46a5460c325", "s4732a14656e310", "sa4d5e376299ae3" ] }, { "age": 80.0, "six": 49.3625, "five": 50.814, "tasks": { "blimp_filtered": 69.78, "supplement_filtered": 60.5, "ewok_filtered": 51.15, "entity_tracking": 20.35, "comps": 52.29, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 58.0 }, "checkpoint": "1d699057f6b267d072dcb65ae929cb3a5835c99003d4b1247c5e9a61457427f7", "precision": "fp32", "verification": "primary", "sources": [ "s5a875c11eb687d", "s985d22b67a2003", "s4732a14656e310", "sa4d5e376299ae3" ] }, { "age": 90.0, "six": 47.665, "five": 50.456, "tasks": { "blimp_filtered": 69.49, "supplement_filtered": 57.41, "ewok_filtered": 51.56, "entity_tracking": 21.56, "comps": 52.26, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 48.0 }, "checkpoint": "3a508dd317fff728f8ebbbb652f926ee44b94adffe61145fd5605666030a5a86", "precision": "fp32", "verification": "primary", "sources": [ "s2a3ca8d27934c9", "s571f7031aa14f6", "s4732a14656e310", "sa4d5e376299ae3" ] }, { "age": 100.0, "six": 48.899166666666666, "five": 50.466, "tasks": { "blimp_filtered": 69.87, "supplement_filtered": 58.78, "ewok_filtered": 50.23, "entity_tracking": 21.22, "comps": 52.23, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 53.0 }, "checkpoint": "0440c3d3f0c72bd801e2c85d466606582a761c0b7ebb734392558944ab12c4db", "precision": "fp32", "verification": "primary", "sources": [ "se7909664f456b6", "s53f3b10cb15634", "s4732a14656e310", "sa4d5e376299ae3" ] } ], "peak": { "five": { "score": 50.814, "ages": [ 80.0 ], "rank": 75 }, "six": { "score": 49.3625, "ages": [ 80.0 ], "rank": 11 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 10, "id": "baby/vistarc_q3_mos2_last3_width384_lr765_20261009", "original_name": "384・MoS2・3層混合・LR765", "title": "384幅・FFN1344・9層", "topology": "MoS2・共通3層混合", "recipe": "主要0.0004667／Embedding0.0007", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1344, "depth": 9, "parts": { "width": 384, "ffn": 1344, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS2", "reference": "7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 2, "reference_mode": "共通で混合", "reference_layers": [ [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00046666666666666666, 0.00039999999999999996, 0.0003333333333333333 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1344, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos2_last3", "attention_received_gate": true, "output_components": 2, "output_state_layers": 3, "reference_attention_width": 256, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym primary31 suites", "original PIQA validation1838" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 exact recurrent prefix graph; checkpoint matched MoS2 last3; TF32off; no changes to probability scoring", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": [ "five axes without PIQA", "SyntaxGym31", "original PIQA acc/acc_norm" ], "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.715, "comparison_scope": "256MoS2/3/4 last3 and384MoS3 last2; external fixed80M peers; own50-100M trajectory" }, "milestones_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "external_protocol": "source:sd1eef500b3fff3" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 40, "sources": [ "s7a5f28c73c8318", "s9da0008c5c347d", "sd9a1b2313330ef", "s593254d462a1d6", "se11351bb009880", "s1d073884a55270", "sd47fe0ce3b3152" ], "points": [ { "age": 50, "six": 48.733333333333334, "five": 50.864, "tasks": { "blimp_filtered": 69.71, "supplement_filtered": 60.17, "ewok_filtered": 51.25, "entity_tracking": 20.94, "comps": 52.25, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 48.0 }, "checkpoint": "0c8ad4c6890527096d32e2dce503ce04b2f71694327933e376ca27fceac246c5", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "scb348443ee527c", "se5e17fae5f0919", "s6fe99760b57908", "sb3a3b58e4488bc", "s6f83623f101f0c", "s790903b86f2161", "s724b0931ad566b", "s7772ecccbb9e1b" ] }, { "age": 60, "six": 48.16166666666667, "five": 50.176, "tasks": { "blimp_filtered": 68.72, "supplement_filtered": 57.87, "ewok_filtered": 51.46, "entity_tracking": 20.86, "comps": 51.97, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 49.0 }, "checkpoint": "f6c7f5d0301ddd83554f102df5f7ec6a46366b6cf22bc908cec385183be025da", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sb73c0e9eba4f7d", "sf394a2badbda51", "s965df61a2bdf9e", "sb3a3b58e4488bc", "s6f83623f101f0c", "s790903b86f2161", "s724b0931ad566b", "s7772ecccbb9e1b" ] }, { "age": 70, "six": 49.3375, "five": 50.992, "tasks": { "blimp_filtered": 70.04, "supplement_filtered": 60.95, "ewok_filtered": 51.93, "entity_tracking": 20.03, "comps": 52.01, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 53.0 }, "checkpoint": "c1af086fe725824fce4c71deb3492b6c02fc980c8a22e83dd85c22b5601c9607", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s25eecf6c6e02ad", "s70300e4b95ffb1", "s400d1d3e4ad3d5", "sb3a3b58e4488bc", "s6f83623f101f0c", "s790903b86f2161", "s724b0931ad566b", "s7772ecccbb9e1b" ] }, { "age": 80, "six": 48.88666666666666, "five": 51.04, "tasks": { "blimp_filtered": 69.08, "supplement_filtered": 60.84, "ewok_filtered": 53.5, "entity_tracking": 19.65, "comps": 52.13, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 51.0 }, "checkpoint": "d97bdbe3a2e0d0e3f805dc335ca7a79f70a682c95a41e39df1e43d0d1d6bc1dc", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5d43312a82399c", "sa55ad3950755ba", "sd15ed25c32da84", "sb3a3b58e4488bc", "s6f83623f101f0c", "s790903b86f2161", "s724b0931ad566b", "s7772ecccbb9e1b" ] }, { "age": 90, "six": 48.545, "five": 50.624, "tasks": { "blimp_filtered": 69.32, "supplement_filtered": 59.52, "ewok_filtered": 52.06, "entity_tracking": 19.93, "comps": 52.29, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 53.0 }, "checkpoint": "236d7aaaa6c5c717e070bf00c106eaa08214aeba44b35404c8e23038d3d4affd", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sf4ce4fed38e8d0", "s1e0098b09d0798", "s5b2f881da95515", "sb3a3b58e4488bc", "s6f83623f101f0c", "s790903b86f2161", "s724b0931ad566b", "s7772ecccbb9e1b" ] }, { "age": 100, "six": 48.410833333333336, "five": 50.96, "tasks": { "blimp_filtered": 68.65, "supplement_filtered": 61.94, "ewok_filtered": 51.53, "entity_tracking": 20.39, "comps": 52.29, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 49.0 }, "checkpoint": "3086fb65134717d95f6174a8e02d4da75f6e12ff64dbd5916a6320bfa7d74a4e", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s9fad63c1a3449c", "s6d979ffa858041", "sbdf2ed17ba9796", "sb3a3b58e4488bc", "s6f83623f101f0c", "s790903b86f2161", "s724b0931ad566b", "s7772ecccbb9e1b" ] } ], "peak": { "five": { "score": 51.04, "ages": [ 80 ], "rank": 42 }, "six": { "score": 49.3375, "ages": [ 70 ], "rank": 12 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 11, "id": "baby/vistarc_q3_mos3_last4_received_value_gate_20261006", "original_name": "256・MoS3・4層混合", "title": "256幅・FFN896・9層", "topology": "MoS3・共通4層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "6・7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 6, 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last4", "attention_received_gate": true, "output_components": 3, "output_state_layers": 4, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3_LAST4 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 80000000, 100000000 ], "objective": { "metric": "maximum observed local six-axis", "train_words": 100000000, "save_words": [ 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "benchmark_words": [ 80000000, 100000000 ], "timing": "after100M training completes; serial", "historical_record": 49.03583333333333, "no_endpoint_penalty": true, "diagnostics": [ "same-checkpoint five-axis and seven component scores", "training loss", "candidate and reader usage", "mixture usage", "runtime" ], "comparison_scope": "MoS3 versus received-gate+MoS2; anticipated80/100M only after owner training instruction and GPU acceptance" } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 53, "sources": [ "sc17f0782fbe42e", "s0b12d86f4417b5", "s1dcebf83340906", "s4a630d1df3d484", "s7d8cc9b5bca58a", "s76be9b90aba4c7", "saafafd5f32a4c6" ], "points": [ { "age": 70, "six": 47.954166666666666, "five": 50.212, "tasks": { "blimp_filtered": 69.15, "supplement_filtered": 58.72, "ewok_filtered": 51.26, "entity_tracking": 19.55, "comps": 52.38, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 51.0 }, "checkpoint": "ec4fefd14ebf158a19674ca17d80c0843c83fdbe888b43e68283ee7ba53f4e24", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s72c43f3b08e1e1", "s15451ddcc64e49", "s9f485487386aab", "s31bd596fb1b632", "s7348bf8e6a478b", "sa8299f0f6db6a7", "s13509749b707d2" ] }, { "age": 80, "six": 49.315, "five": 51.054, "tasks": { "blimp_filtered": 70.0, "supplement_filtered": 61.9, "ewok_filtered": 51.65, "entity_tracking": 19.29, "comps": 52.43, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 56.0 }, "checkpoint": "e97e3281516b83c549d8fe892e0ee6238ed38e9fcfbc5817a8b7d13fbb48d2ab", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s773883ce0051b4", "s15451ddcc64e49", "s60098a3de766e3", "s31bd596fb1b632", "s7348bf8e6a478b", "sa8299f0f6db6a7", "s13509749b707d2" ] }, { "age": 90, "six": 47.605, "five": 49.902, "tasks": { "blimp_filtered": 68.64, "supplement_filtered": 59.67, "ewok_filtered": 49.88, "entity_tracking": 19.34, "comps": 51.98, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 47.0 }, "checkpoint": "6a599b9c1113977e52185135cb142539a52c3b936bdf95c816bc4aeb18360caf", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s2a344ab229fdac", "s15451ddcc64e49", "s9ae20735a7e6a8", "s31bd596fb1b632", "s7348bf8e6a478b", "sa8299f0f6db6a7", "s13509749b707d2" ] }, { "age": 100, "six": 47.87583333333333, "five": 50.33, "tasks": { "blimp_filtered": 68.34, "supplement_filtered": 58.67, "ewok_filtered": 53.1, "entity_tracking": 19.12, "comps": 52.42, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 45.0 }, "checkpoint": "3b602314b745f02401ed1219ee8f223a3196e2f7f846f3e40e760a4521491298", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s853f3d0c8bfb48", "s15451ddcc64e49", "s182224115425c7", "s31bd596fb1b632", "s7348bf8e6a478b", "sa8299f0f6db6a7", "s13509749b707d2" ] } ], "peak": { "five": { "score": 51.054, "ages": [ 80 ], "rank": 40 }, "six": { "score": 49.315, "ages": [ 80 ], "rank": 13 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 12, "id": "run/IPRLM320_LR070605_MAIN_MATRICES_20261012", "original_name": "IPRLM320_LR070605_MAIN_MATRICES_20261012", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at70/80/90/100M", "SyntaxGym31 and originalPIQA1838 at70/80/90/100M", "frozen bAbI1947 and ProofWriter1000 at70/80/90/100M" ], "intermediate": "No evaluation during training; all benchmarks after100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Original parent320 and child matched70/80/90/100M, with observed peaks separately" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 71, "sources": [ "s939a71b7ec415f", "s8c18e536dac4be", "s9f94e8e4fb7d05", "sf02275f2c4d1d2", "sac7cfeee4fd116", "s3ee3261d657130", "s90df22b0f38de2", "s363bff5308fb34", "sf9cc1dd7ba351a", "s08373efbe779ce", "s807ae79332bc70", "sf22f093ef2ddbd" ], "points": [ { "age": 40.0, "six": 47.9375, "five": 50.58, "tasks": { "blimp_filtered": 68.77, "supplement_filtered": 58.24, "ewok_filtered": 51.2, "entity_tracking": 21.68, "comps": 53.01, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 51.0 }, "checkpoint": "186fc52623465cd1987c98261feb2311d543458d6329c75d13e128bd4a2ca394", "precision": "fp32", "verification": "primary", "sources": [ "sec6933ec7d9601", "s3a08af141e244d" ] }, { "age": 70.0, "six": 48.611666666666665, "five": 50.998, "tasks": { "blimp_filtered": 69.45, "supplement_filtered": 60.34, "ewok_filtered": 53.13, "entity_tracking": 19.47, "comps": 52.6, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 52.0 }, "checkpoint": "5544cb8631a316a837181a3e66ea9cc7b72cc18213f1689d654ccaa489bb772e", "precision": "fp32", "verification": "primary", "sources": [ "s0cf394dd13b70a", "s319360c884580f", "se6f9c820d9ac11", "s2ec94ab325cad2", "sad0779c464106b", "sacc2001f2887af" ] }, { "age": 80.0, "six": 49.2875, "five": 51.406, "tasks": { "blimp_filtered": 69.71, "supplement_filtered": 63.74, "ewok_filtered": 51.95, "entity_tracking": 19.05, "comps": 52.58, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 57.0 }, "checkpoint": "c92e74f5873b2d6a1830a14ae3dada672fe4c140ad51c3f5be088e0430685c3f", "precision": "fp32", "verification": "primary", "sources": [ "sf0ed7bd2898bf5", "s036ec2df4aca50", "se6f9c820d9ac11", "s2ec94ab325cad2", "sad0779c464106b", "sacc2001f2887af" ] }, { "age": 90.0, "six": 48.8325, "five": 50.872, "tasks": { "blimp_filtered": 68.56, "supplement_filtered": 60.46, "ewok_filtered": 52.18, "entity_tracking": 20.55, "comps": 52.61, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 53.0 }, "checkpoint": "e40c7c090c91d2edd640bb43bbb545c89be40714ef4a526744eb75d22287762c", "precision": "fp32", "verification": "primary", "sources": [ "s9f79c2157450f0", "s4550fedcc58827", "se6f9c820d9ac11", "s2ec94ab325cad2", "sad0779c464106b", "sacc2001f2887af" ] }, { "age": 100.0, "six": 48.5575, "five": 51.036, "tasks": { "blimp_filtered": 68.55, "supplement_filtered": 61.88, "ewok_filtered": 50.34, "entity_tracking": 21.55, "comps": 52.86, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 50.0 }, "checkpoint": "81b0b7ef24596567aa8ed19d0657f07d9f85c15fd068c3b6ccf059186cf5abe5", "precision": "fp32", "verification": "primary", "sources": [ "s7afa73dcdfa7d5", "s8b0cce43c4202f", "se6f9c820d9ac11", "s2ec94ab325cad2", "sad0779c464106b", "sacc2001f2887af" ] } ], "peak": { "five": { "score": 51.406, "ages": [ 80.0 ], "rank": 11 }, "six": { "score": 49.2875, "ages": [ 80.0 ], "rank": 14 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 13, "id": "baby/vistarc_q3_mos3_received_value_gate_20261005", "original_name": "256・MoS3・3層混合", "title": "256幅・FFN896・9層", "topology": "MoS3", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last3", "attention_received_gate": true, "output_components": 3, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 80000000, 100000000 ], "objective": { "metric": "maximum observed local six-axis", "train_words": 100000000, "save_words": [ 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "benchmark_words": [ 80000000, 100000000 ], "timing": "after100M training completes; serial", "historical_record": 49.03583333333333, "no_endpoint_penalty": true, "diagnostics": [ "same-checkpoint five-axis and seven component scores", "training loss", "candidate and reader usage", "mixture usage", "runtime" ], "comparison_scope": "MoS3 versus received-gate+MoS2; anticipated80/100M only after owner training instruction and GPU acceptance" } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 57, "sources": [ "s4741b5d795f4c5", "s57d3de66b940a4", "s3c3eabcbcb443d", "s26da1c468ab645", "s3587426dcca58f", "sa77f2308ee9648", "sa3c7912a35618b", "s52de481896c42b", "s2f4f5962abdbb4", "se0ed1712cc64f0", "se55226b9ff72ca" ], "points": [ { "age": 40.0, "six": 48.5025, "five": 50.676, "tasks": { "blimp_filtered": 68.7, "supplement_filtered": 61.04, "ewok_filtered": 52.27, "entity_tracking": 19.35, "comps": 52.02, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 51.0 }, "checkpoint": "0629f004c53b3283ca824d9d6bcab98665e976a271cf9f0c31ef3de5b86873f5", "precision": "fp32", "verification": "primary", "sources": [ "se67b02f7dcb3ef", "s3284cf5480c65b" ] }, { "age": 70, "six": 48.225, "five": 50.746, "tasks": { "blimp_filtered": 69.93, "supplement_filtered": 59.51, "ewok_filtered": 51.58, "entity_tracking": 20.83, "comps": 51.88, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 46.0 }, "checkpoint": "bf5aadc05a1b59176cd5bc321f478f72571cd0c6de579112d4e21442235e56c5", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sabd0bde8ebbcfe", "scc90fadf5630dd" ] }, { "age": 80, "six": 49.2675, "five": 51.488, "tasks": { "blimp_filtered": 69.26, "supplement_filtered": 62.98, "ewok_filtered": 51.44, "entity_tracking": 21.41, "comps": 52.35, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 54.0 }, "checkpoint": "8346ea666299a7db42137de508c07cc0d8684ee092551fc315afbc4e90d34139", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s34ccc8ab603a38", "s49e2439f68186d", "s50f7cb39f8470d", "s4fe749f4b2f7d0", "scc6b00f7daef02", "sb054506eb10357", "sba487c6cb5846d", "s0c010113d093bd", "sd7ec99d3a7dbbe" ] }, { "age": 90, "six": 47.986666666666665, "five": 50.154, "tasks": { "blimp_filtered": 68.61, "supplement_filtered": 58.46, "ewok_filtered": 51.61, "entity_tracking": 19.89, "comps": 52.2, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "759ae407d1e52d3d718ffdeaaa8dfbe04e4637daa01226e646de79ad0f99dcb4", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sb0c138aa2e2e61", "se93ea242467bba" ] }, { "age": 100, "six": 47.718333333333334, "five": 50.52, "tasks": { "blimp_filtered": 67.98, "supplement_filtered": 59.14, "ewok_filtered": 52.0, "entity_tracking": 21.0, "comps": 52.48, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 48.0 }, "checkpoint": "f74648f4c7eb212c4035d13c2e7c969648675fe09eff5548d69915dd0edf458a", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s794cdee1d455cc", "s49e2439f68186d", "s50f7cb39f8470d", "s61cf89a1c8f367", "scc6b00f7daef02", "sb054506eb10357", "sba487c6cb5846d", "s0c010113d093bd" ] } ], "peak": { "five": { "score": 51.488, "ages": [ 80 ], "rank": 8 }, "six": { "score": 49.2675, "ages": [ 80 ], "rank": 15 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 14, "id": "baby/vistarc_q3_mos4_last3_width384_20261010", "original_name": "今回384・MoS4・3層混合", "title": "384幅・FFN1344・9層", "topology": "MoS4・共通3層混合", "recipe": "主要0.0005333／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1344, "depth": 9, "parts": { "width": 384, "ffn": 1344, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS4", "reference": "7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 4, "reference_mode": "共通で混合", "reference_layers": [ [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.0005333333333333334, 0.00046666666666666666, 0.00039999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1344, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos4_last3", "attention_received_gate": true, "output_components": 4, "output_state_layers": 3, "reference_attention_width": 256, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "PIQA1838 acc and acc_norm", "bAbI1947 and ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 exact recurrent prefix graph; checkpoint matched MoS4 last3; TF32off; unchanged probability scoring", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": [ "five axes without PIQA", "SyntaxGym31", "original PIQA acc/acc_norm" ], "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.8675, "comparison_scope": "256MoS4last3;384MoS3last2;320MoS3last2 at70M external;320branch123 at80M local peak" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ], "external_protocol": "source:sd1eef500b3fff3", "grounded_words": [ 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 59, "sources": [ "s256fe6181b5f6c", "s11cf12ba58d7fc", "s5e5bfee59cc937", "s9666d5e4ce0bc7", "s5f400843fc58b7", "sfc55b415bfbf0c", "s86a1d7e66a0411" ], "points": [ { "age": 70, "six": 48.435, "five": 50.598, "tasks": { "blimp_filtered": 70.08, "supplement_filtered": 57.97, "ewok_filtered": 51.83, "entity_tracking": 20.7, "comps": 52.41, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 50.0 }, "checkpoint": "cb97122322a5a6c9d5f0826b4cef098e2fef975db8b149b9ebe0df515a3b2ca1", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sef68b7e7aa22d6", "s41cab18d11b4a1", "se127884ad83953", "s70786401960f88", "s68060d28b54c39", "s4c3b845c644a50", "s17eae876b5bb73", "s8efc420fb80762" ] }, { "age": 80, "six": 49.25833333333333, "five": 51.086, "tasks": { "blimp_filtered": 69.88, "supplement_filtered": 61.75, "ewok_filtered": 51.62, "entity_tracking": 20.0, "comps": 52.18, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 55.0 }, "checkpoint": "a789d081f51ab1fc368362aefab2b50cbc0237fd041ce7f1ee537b5627b36827", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sb1e570ffaad073", "sa9ffc043c54240", "scf6bfd12cfcf75", "s70786401960f88", "s68060d28b54c39", "s4c3b845c644a50", "s17eae876b5bb73", "s8efc420fb80762" ] }, { "age": 90, "six": 48.085, "five": 50.566, "tasks": { "blimp_filtered": 69.82, "supplement_filtered": 58.53, "ewok_filtered": 52.36, "entity_tracking": 20.13, "comps": 51.99, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 50.0 }, "checkpoint": "bb2d7460165118e13fc4665ecee36c27233bec9ff0cac77d577fd3096fba81ae", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sdc697c34cc5b14", "scd28a166e256de", "se9c70cf589e8d3", "s70786401960f88", "s68060d28b54c39", "s4c3b845c644a50", "s17eae876b5bb73", "s8efc420fb80762" ] }, { "age": 100, "six": 48.089166666666664, "five": 50.386, "tasks": { "blimp_filtered": 69.09, "supplement_filtered": 57.36, "ewok_filtered": 51.79, "entity_tracking": 21.51, "comps": 52.18, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 47.0 }, "checkpoint": "5ddf39b3852de0944c19b525e41016af031cd82b6f27202d88d51c8e9b814d1d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sa78209b0fde558", "s64dba3218d313e", "scf7dd1903a67c2", "s70786401960f88", "s68060d28b54c39", "s4c3b845c644a50", "s17eae876b5bb73", "s8efc420fb80762" ] } ], "peak": { "five": { "score": 51.086, "ages": [ 80 ], "rank": 37 }, "six": { "score": 49.25833333333333, "ages": [ 80 ], "rank": 16 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 15, "id": "run/IPRLM320_DIRECT_ROUTER_20261012", "original_name": "IPRLM320_DIRECT_ROUTER_20261012", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "direct_supervised_v1", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123_direct_router", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "output_mixture_gate": "direct_supervised_v1", "mixture_gate_hidden": 320, "router_norm_eps": 1e-06, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.715, "comparison_scope": "exact320 branch123 parent", "parent_peak_six_axis": 49.8675 }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ], "authorized": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 73, "sources": [ "s719c7a7d20457f", "s321a9818b979fb", "sf4307bcc9f41f7", "s47c2da9c03aece", "sad76f3a06b9a12", "s914ad588e42f55" ], "points": [ { "age": 70.0, "six": 49.06583333333333, "five": 50.958, "tasks": { "blimp_filtered": 69.71, "supplement_filtered": 59.71, "ewok_filtered": 52.17, "entity_tracking": 20.84, "comps": 52.36, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 53.0 }, "checkpoint": "7a6a0da5b62aa1871a51533ffab065b7a902feb1df7db3769d5c91f2e352e9f9", "precision": "fp32", "verification": "primary", "sources": [ "s3b8aa9d84ecb48", "sf80ecd0fb07ce8", "s42302d6ed0cc26", "sf72d5a49762d09", "sd287039deffda7", "s5cb0dbb480e05b" ] }, { "age": 80.0, "six": 49.208333333333336, "five": 50.832, "tasks": { "blimp_filtered": 68.97, "supplement_filtered": 60.94, "ewok_filtered": 52.22, "entity_tracking": 20.03, "comps": 52.0, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 55.0 }, "checkpoint": "988318b47c11fb10e6edaa177194f3d868d3499ee196302512cc2ea5b4a865af", "precision": "fp32", "verification": "primary", "sources": [ "sc8848345c5107b", "sbcd843ce73a8b9", "s42302d6ed0cc26", "sf72d5a49762d09", "sd287039deffda7", "s5cb0dbb480e05b" ] }, { "age": 90.0, "six": 48.026666666666664, "five": 50.114, "tasks": { "blimp_filtered": 68.0, "supplement_filtered": 59.94, "ewok_filtered": 49.78, "entity_tracking": 21.23, "comps": 51.62, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 48.0 }, "checkpoint": "f4e705625d8bd8c1c77296cd3f211cd6aea490934d9eda01a49ed7b936ff3d3b", "precision": "fp32", "verification": "primary", "sources": [ "s6fe7c90d747bdd", "s12a7ae39b02d5c", "s42302d6ed0cc26", "sf72d5a49762d09", "sd287039deffda7", "s5cb0dbb480e05b" ] }, { "age": 100.0, "six": 48.5625, "five": 50.648, "tasks": { "blimp_filtered": 68.58, "supplement_filtered": 58.53, "ewok_filtered": 52.45, "entity_tracking": 21.51, "comps": 52.17, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 52.0 }, "checkpoint": "1f95cd171e47c001e8d4b7c4e7fca16118ce21b39b56c6eb88074dd252a460b5", "precision": "fp32", "verification": "primary", "sources": [ "s58aeab63d94c57", "s6498f97ff298ff", "s42302d6ed0cc26", "sf72d5a49762d09", "sd287039deffda7", "s5cb0dbb480e05b" ] } ], "peak": { "five": { "score": 50.958, "ages": [ 70.0 ], "rank": 49 }, "six": { "score": 49.208333333333336, "ages": [ 80.0 ], "rank": 18 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 16, "id": "run/IPRLM320_FFN_ADDITIVE_BYPASS_SEED731946281_20261014", "original_name": "IPRLM320_FFN_ADDITIVE_BYPASS_SEED731946281_20261014", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After each40M training; no automatic extension.", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M training", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Both matchedseed parent320 at40M;40M allocates budget, does not establish final peak." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 79, "sources": [ "s14b8d5e0b1b234", "sfc0a2499c09697", "s9695d8c828d172" ], "points": [ { "age": 40.0, "six": 47.50416666666667, "five": 49.978, "tasks": { "blimp_filtered": 67.72, "supplement_filtered": 58.42, "ewok_filtered": 51.88, "entity_tracking": 19.97, "comps": 51.9, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 46.0 }, "checkpoint": "c5d96991400965b4435b9c9955d826f86bd472b8a258753cf911873cc3e7681f", "precision": "fp32", "verification": "primary", "sources": [ "s341472ac808ce1", "s4bcd08e865243d" ] }, { "age": 70.0, "six": 48.589166666666664, "five": 50.974, "tasks": { "blimp_filtered": 69.59, "supplement_filtered": 59.06, "ewok_filtered": 51.67, "entity_tracking": 22.32, "comps": 52.23, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 51.0 }, "checkpoint": "51fc2284e6c37c05040f0fa92e47c4fd6135a070d82248552f78116d489cbb56", "precision": "fp32", "verification": "primary", "sources": [ "sdcf28e6b4a24f1", "s39922da907e770" ] }, { "age": 80.0, "six": 49.205, "five": 51.016, "tasks": { "blimp_filtered": 68.41, "supplement_filtered": 60.44, "ewok_filtered": 52.03, "entity_tracking": 21.91, "comps": 52.29, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 57.0 }, "checkpoint": "72e9f6602de565e4f92d7d81090fceb693b8578aae2268957e0c0d3e8906f0f0", "precision": "fp32", "verification": "primary", "sources": [ "s688f763ffd538e", "sd5ef3b8bc63260" ] } ], "peak": { "five": { "score": 51.016, "ages": [ 80.0 ], "rank": 43 }, "six": { "score": 49.205, "ages": [ 80.0 ], "rank": 19 } }, "clear_points": [ { "number": 16, "age": 80, "seed": 731946281, "checkpoint_sha256": "72e9f6602de565e4f92d7d81090fceb693b8578aae2268957e0c0d3e8906f0f0", "old_six": 49.205, "old_five": 51.016, "old_tasks": { "blimp_filtered": 68.41, "supplement_filtered": 60.44, "ewok_filtered": 52.03, "entity_tracking": 21.91, "comps": 52.29, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 57.0 }, "correct": 12246, "total": 17274, "accuracy": 70.89267106634249, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6896, "total": 8392, "ties": 0 }, "comps": { "correct": 3899, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 4, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 13, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1038, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 2, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 2, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 2, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 2, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 77, "total": 114, "ties": 0 }, "boolq": { "correct": 10, "total": 53, "ties": 0 }, "mnli": { "correct": 35, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 65.40347600800305, "checkpoint_rank": 9, "rank": 8 } ], "clear_peak": { "number": 16, "age": 80, "seed": 731946281, "checkpoint_sha256": "72e9f6602de565e4f92d7d81090fceb693b8578aae2268957e0c0d3e8906f0f0", "old_six": 49.205, "old_five": 51.016, "old_tasks": { "blimp_filtered": 68.41, "supplement_filtered": 60.44, "ewok_filtered": 52.03, "entity_tracking": 21.91, "comps": 52.29, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 57.0 }, "correct": 12246, "total": 17274, "accuracy": 70.89267106634249, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6896, "total": 8392, "ties": 0 }, "comps": { "correct": 3899, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 4, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 13, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1038, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 2, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 2, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 2, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 2, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 77, "total": 114, "ties": 0 }, "boolq": { "correct": 10, "total": 53, "ties": 0 }, "mnli": { "correct": 35, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 65.40347600800305, "checkpoint_rank": 9, "rank": 8 }, "present": true }, { "number": 17, "id": "run/IPRLM_MOS3_320_LAST2_SEED731946281_100M_20261014", "original_name": "IPRLM_MOS3_320_LAST2_SEED731946281_100M_20261014", "title": "320幅・FFN1120・9層", "topology": "MoS3・共通2層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last2", "attention_received_gate": true, "output_components": 3, "output_state_layers": 2, "reference_attention_width": 256, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes" ], "intermediate": "After100M train:70/80/90/100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "selection_metric": "five_axis", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "no_endpoint_penalty": true }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 153, "sources": [ "s8b40c168add0c2", "sc750b331c008f3", "s86dc8f7b2ee202" ], "points": [ { "age": 70.0, "six": 48.060833333333335, "five": 50.634, "tasks": { "blimp_filtered": 69.09, "supplement_filtered": 58.84, "ewok_filtered": 51.95, "entity_tracking": 21.08, "comps": 52.21, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 50.0 }, "checkpoint": "7b175927c60f697c631a4a7bcbf8483cb34030206fd8f5eede6f7b76c4ed9774", "precision": "fp32", "verification": "primary", "sources": [ "s88ef890036ff05", "s0ca68b84c9a335" ] }, { "age": 80.0, "six": 48.9025, "five": 51.35, "tasks": { "blimp_filtered": 68.61, "supplement_filtered": 62.56, "ewok_filtered": 52.66, "entity_tracking": 20.09, "comps": 52.83, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 51.0 }, "checkpoint": "77e4dfba11f567eaefffba3d8f9d592879c2939c99e2c9cd53aa816ebc3edda8", "precision": "fp32", "verification": "primary", "sources": [ "s78812b313046de", "saa99e285c03e27" ] }, { "age": 90.0, "six": 48.26833333333333, "five": 50.206, "tasks": { "blimp_filtered": 67.08, "supplement_filtered": 59.01, "ewok_filtered": 52.18, "entity_tracking": 20.25, "comps": 52.51, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 49.0 }, "checkpoint": "3509fba97ada1e08685674e73d0722f91eb31e9000308c825069cd322be142ec", "precision": "fp32", "verification": "primary", "sources": [ "s4c91a97e708e7b", "s931f94f756baa3" ] }, { "age": 100.0, "six": 49.16166666666667, "five": 50.776, "tasks": { "blimp_filtered": 69.36, "supplement_filtered": 59.46, "ewok_filtered": 52.35, "entity_tracking": 20.29, "comps": 52.42, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 55.0 }, "checkpoint": "1fde136a4a0c858d3df0fff6521bc4cf8db6297e61bcc48245e6af8e7d30f168", "precision": "fp32", "verification": "primary", "sources": [ "s736be91f1d68a8", "sb70228d810dbf9" ] } ], "peak": { "five": { "score": 51.35, "ages": [ 80.0 ], "rank": 17 }, "six": { "score": 49.16166666666667, "ages": [ 100.0 ], "rank": 20 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 18, "id": "run/IPRLM384_READOUT456_PREDICTION504_MIXING480_SEED20260907_80M_20261023", "original_name": "IPRLM384_READOUT456_PREDICTION504_MIXING480_SEED20260907_80M_20261023", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "層読出し456／予測504／混合480", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40/70/80M after80M training" ], "intermediate": "saved40/70M states evaluated after80M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000, 70000000, 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000, 70000000, 80000000 ], "timing": "after80M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All29LRconditions plusparent320/fullinventory; same-age and observed peaks" }, "milestones_words": [ 40000000, 70000000, 80000000 ], "checkpoints_words": [ 40000000, 70000000, 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 142, "sources": [ "s0c4212d34a1da0", "s8422216e860cca" ], "points": [ { "age": 40.0, "six": 48.211666666666666, "five": 50.912, "tasks": { "blimp_filtered": 70.38, "supplement_filtered": 60.66, "ewok_filtered": 51.15, "entity_tracking": 19.79, "comps": 52.58, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 50.0 }, "checkpoint": "48a07603b355cf926ebc2229ff9dff9ad4a5d19201268347304567737f965472", "precision": "fp32", "verification": "primary", "sources": [ "s3d4776ae968b43", "sf9ce3fa4897ece" ] }, { "age": 70.0, "six": 48.681666666666665, "five": 50.982, "tasks": { "blimp_filtered": 70.54, "supplement_filtered": 60.11, "ewok_filtered": 51.97, "entity_tracking": 19.56, "comps": 52.73, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 53.0 }, "checkpoint": "c611ba537ae4d4e44d4af1f7ba233ad6c664ab7fdae8cb49604e4a057c269fa5", "precision": "fp32", "verification": "primary", "sources": [ "sa01e009bafcb41", "sf9ce3fa4897ece" ] }, { "age": 80.0, "six": 49.14333333333333, "five": 51.142, "tasks": { "blimp_filtered": 69.08, "supplement_filtered": 61.14, "ewok_filtered": 52.14, "entity_tracking": 20.91, "comps": 52.44, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 55.0 }, "checkpoint": "9264d9b9089338c59451ac9ad33b1d4e5900e587556ce674c0acdb92007fda5a", "precision": "fp32", "verification": "primary", "sources": [ "s52d9b5c9db8242", "sf9ce3fa4897ece" ] } ], "peak": { "five": { "score": 51.142, "ages": [ 80.0 ], "rank": 29 }, "six": { "score": 49.14333333333333, "ages": [ 80.0 ], "rank": 21 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 19, "id": "run/IPRLM336_FFN1232_ADDITIVE_SEED20260907_20261016", "original_name": "IPRLM336_FFN1232_ADDITIVE_SEED20260907_20261016", "title": "336幅・FFN1232・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M parent320/additive320/additive336FFN1176 and siblingFFN variants; observed later peaks separate." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 91, "sources": [ "s428cef8dd0a90c", "s71ee15f48f81c4", "sb80244a946ad3f" ], "points": [ { "age": 40.0, "six": 48.266666666666666, "five": 50.69, "tasks": { "blimp_filtered": 70.29, "supplement_filtered": 58.86, "ewok_filtered": 52.35, "entity_tracking": 19.36, "comps": 52.59, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 49.0 }, "checkpoint": "9da42e5181b1146c369af761402d5c001a481c143dd9cf99db344205c3641a5b", "precision": "fp32", "verification": "primary", "sources": [ "s8948447779c000", "s61069d633f5e42" ] }, { "age": 70.0, "six": 49.14083333333333, "five": 51.248, "tasks": { "blimp_filtered": 69.75, "supplement_filtered": 60.94, "ewok_filtered": 52.6, "entity_tracking": 20.44, "comps": 52.51, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 51.0 }, "checkpoint": "9bc8109d0cb789771079b5a9708a1e10be831da6f1e217a6d92b46d825dc81a3", "precision": "fp32", "verification": "primary", "sources": [ "sb4fc87dbf3cd6a", "s33d2bbe8ea0d98" ] }, { "age": 80.0, "six": 49.07083333333333, "five": 50.972, "tasks": { "blimp_filtered": 68.9, "supplement_filtered": 61.62, "ewok_filtered": 51.7, "entity_tracking": 20.28, "comps": 52.36, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 50.0 }, "checkpoint": "0a3bffc09cf7acd361bb1a8f8637c9bf472338a4ed3441a3aebba5eb5b1cfe21", "precision": "fp32", "verification": "primary", "sources": [ "s8cbd397fb8108e", "s1c5c0191b8ed88" ] } ], "peak": { "five": { "score": 51.248, "ages": [ 70.0 ], "rank": 23 }, "six": { "score": 49.14083333333333, "ages": [ 70.0 ], "rank": 22 } }, "clear_points": [ { "number": 19, "age": 70, "seed": 20260907, "checkpoint_sha256": "9bc8109d0cb789771079b5a9708a1e10be831da6f1e217a6d92b46d825dc81a3", "old_six": 49.14083333333333, "old_five": 51.248, "old_tasks": { "blimp_filtered": 69.75, "supplement_filtered": 60.94, "ewok_filtered": 52.6, "entity_tracking": 20.44, "comps": 52.51, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 51.0 }, "correct": 12493, "total": 17274, "accuracy": 72.32256570568485, "ties": 0, "tasks": { "blimp_filtered": { "correct": 7031, "total": 8392, "ties": 0 }, "comps": { "correct": 3921, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 14, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1026, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 74, "total": 114, "ties": 0 }, "boolq": { "correct": 23, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 310, "total": 947, "ties": 0 }, "rte": { "correct": 3, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 63.872348315002455, "checkpoint_rank": 2, "rank": 2 } ], "clear_peak": { "number": 19, "age": 70, "seed": 20260907, "checkpoint_sha256": "9bc8109d0cb789771079b5a9708a1e10be831da6f1e217a6d92b46d825dc81a3", "old_six": 49.14083333333333, "old_five": 51.248, "old_tasks": { "blimp_filtered": 69.75, "supplement_filtered": 60.94, "ewok_filtered": 52.6, "entity_tracking": 20.44, "comps": 52.51, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 51.0 }, "correct": 12493, "total": 17274, "accuracy": 72.32256570568485, "ties": 0, "tasks": { "blimp_filtered": { "correct": 7031, "total": 8392, "ties": 0 }, "comps": { "correct": 3921, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 14, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1026, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 74, "total": 114, "ties": 0 }, "boolq": { "correct": 23, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 310, "total": 947, "ties": 0 }, "rte": { "correct": 3, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 63.872348315002455, "checkpoint_rank": 2, "rank": 2 }, "present": true }, { "number": 20, "id": "baby/vistarc_q3_mos3_last2_received_value_gate_20261006", "original_name": "256・MoS3・2層混合", "title": "256幅・FFN896・9層", "topology": "MoS3・共通2層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last2", "attention_received_gate": true, "output_components": 3, "output_state_layers": 2, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3_LAST2 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 80000000, 100000000 ], "objective": { "metric": "maximum observed local six-axis", "train_words": 100000000, "save_words": [ 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "benchmark_words": [ 80000000, 100000000 ], "timing": "after100M training completes; serial", "historical_record": 49.03583333333333, "no_endpoint_penalty": true, "diagnostics": [ "same-checkpoint five-axis and seven component scores", "training loss", "candidate and reader usage", "mixture usage", "runtime" ], "comparison_scope": "MoS3 versus received-gate+MoS2; anticipated80/100M only after owner training instruction and GPU acceptance" } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 37, "sources": [ "s27822c263b8946", "s101116cf5fad80", "s25f171cb466160", "sf08900b2d12c14", "se1e2cfe86b8f27", "s8761228e256049", "s3fe6d605bc5f66" ], "points": [ { "age": 40.0, "six": 47.9825, "five": 50.052, "tasks": { "blimp_filtered": 69.67, "supplement_filtered": 57.21, "ewok_filtered": 52.09, "entity_tracking": 19.87, "comps": 51.42, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 51.0 }, "checkpoint": "c47cafd558b1a8a0704d5bb9d2f10022c9e031fbd80d70b0d325e70a3cc766e8", "precision": "fp32", "verification": "primary", "sources": [ "s9948dd16790ede", "sf3d3e513a73963" ] }, { "age": 50, "six": 48.21666666666667, "five": 50.63, "tasks": { "blimp_filtered": 69.42, "supplement_filtered": 59.12, "ewok_filtered": 51.25, "entity_tracking": 21.18, "comps": 52.18, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 49.0 }, "checkpoint": "693846974bd237cf160add601716c16736d0022ce9a3404794f22651f3af3c77", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s831c1f3df02a0e" ] }, { "age": 60, "six": 47.560833333333335, "five": 50.234, "tasks": { "blimp_filtered": 67.83, "supplement_filtered": 57.4, "ewok_filtered": 52.78, "entity_tracking": 21.44, "comps": 51.72, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 48.0 }, "checkpoint": "f1338164d9718305d3084340ecb4a4afd0084167d40229a85b5ab7c067e67697", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s9c71a7a0495647" ] }, { "age": 70, "six": 48.0125, "five": 50.482, "tasks": { "blimp_filtered": 70.3, "supplement_filtered": 57.46, "ewok_filtered": 54.05, "entity_tracking": 18.73, "comps": 51.87, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 49.0 }, "checkpoint": "715a445cf5f2ac451834a919c8f0b7053e850a82c6fe46c4d45b81c7b09563ca", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s88c3c4f90b5ca6", "s14c1311f07fdfe", "s758bd3bd9b9003", "s8527272e67f77e", "sffd7904ede87ac", "sb0b3a6ce72fddf", "sc1b05281b645be" ] }, { "age": 80, "six": 49.1375, "five": 51.72, "tasks": { "blimp_filtered": 70.27, "supplement_filtered": 62.8, "ewok_filtered": 52.71, "entity_tracking": 20.64, "comps": 52.18, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 54.0 }, "checkpoint": "7bfcf3c02081e8a549103f15be4a7eb33a20887ec621ca68abd8a50cc7293982", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s077673944a987f", "s14c1311f07fdfe", "sf556fa3a2679d1", "s8527272e67f77e", "sffd7904ede87ac", "sb0b3a6ce72fddf", "sc1b05281b645be" ] }, { "age": 90, "six": 48.45666666666666, "five": 50.524, "tasks": { "blimp_filtered": 70.15, "supplement_filtered": 57.92, "ewok_filtered": 52.73, "entity_tracking": 19.81, "comps": 52.01, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 51.0 }, "checkpoint": "d93cd53fc1b23d8c44e7967d37a894e8510325a5bc8a962fe5128d628e0f674e", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s7c114e045d7484", "s14c1311f07fdfe", "s52c54ded345490", "s8527272e67f77e", "sffd7904ede87ac", "sb0b3a6ce72fddf", "sc1b05281b645be" ] }, { "age": 100, "six": 47.696666666666665, "five": 50.588, "tasks": { "blimp_filtered": 69.09, "supplement_filtered": 59.0, "ewok_filtered": 52.61, "entity_tracking": 19.95, "comps": 52.29, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 49.0 }, "checkpoint": "664915473993bcc1b7d2832668c464181cfedb4cf2823983e840312199e372fe", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sbe21e6255fec3d", "s14c1311f07fdfe", "s20dc876af23282", "s8527272e67f77e", "sffd7904ede87ac", "sb0b3a6ce72fddf", "sc1b05281b645be" ] } ], "peak": { "five": { "score": 51.72, "ages": [ 80 ], "rank": 2 }, "six": { "score": 49.1375, "ages": [ 80 ], "rank": 23 } }, "clear_points": [ { "number": 20, "age": 80, "seed": 20260907, "checkpoint_sha256": "7bfcf3c02081e8a549103f15be4a7eb33a20887ec621ca68abd8a50cc7293982", "old_six": 49.1375, "old_five": 51.72, "old_tasks": { "blimp_filtered": 70.27, "supplement_filtered": 62.8, "ewok_filtered": 52.71, "entity_tracking": 20.64, "comps": 52.18, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 54.0 }, "correct": 12281, "total": 17274, "accuracy": 71.095287715642, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6956, "total": 8392, "ties": 0 }, "comps": { "correct": 3885, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 18, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1026, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 2, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 2, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 3, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 79, "total": 114, "ties": 0 }, "boolq": { "correct": 12, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 216, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 58.81191615100397, "checkpoint_rank": 6, "rank": 6 } ], "clear_peak": { "number": 20, "age": 80, "seed": 20260907, "checkpoint_sha256": "7bfcf3c02081e8a549103f15be4a7eb33a20887ec621ca68abd8a50cc7293982", "old_six": 49.1375, "old_five": 51.72, "old_tasks": { "blimp_filtered": 70.27, "supplement_filtered": 62.8, "ewok_filtered": 52.71, "entity_tracking": 20.64, "comps": 52.18, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 54.0 }, "correct": 12281, "total": 17274, "accuracy": 71.095287715642, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6956, "total": 8392, "ties": 0 }, "comps": { "correct": 3885, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 18, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1026, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 2, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 2, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 3, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 79, "total": 114, "ties": 0 }, "boolq": { "correct": 12, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 216, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 58.81191615100397, "checkpoint_rank": 6, "rank": 6 }, "present": true }, { "number": 21, "id": "run/VISTA336_CANDIDATE_READ_RECEIVED_GATE_MOS3_READER48_SEED20260907_20261018", "original_name": "VISTA336_CANDIDATE_READ_RECEIVED_GATE_MOS3_READER48_SEED20260907_20261018", "title": "336幅・FFN1232・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTA_CANDIDATE_READ_RECEIVED_GATE_MOS3_READER48", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "mos3", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": false, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336, "reader_width": 48, "candidate_variance_scale_init": 0.1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at70M", "local six axes at80M" ], "intermediate": "After80Mtraining, evaluate70M and80M checkpoints", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 TF32off, MoS3 + fullcandidate + Reader48; one live decoder shape; canonical parity verified.", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 70000000, 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000 ], "timing": "after80Mtraining", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same seed20260907 at70/80M; direct newcandidate+received+MoS3 parent and old336 Reader+FFNadd reference." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 162, "sources": [ "s9af5c4f1f60fdc", "s38fd9573e7bd53", "s71854f5682ebd6" ], "points": [ { "age": 70.0, "six": 48.1125, "five": 50.802, "tasks": { "blimp_filtered": 70.09, "supplement_filtered": 59.24, "ewok_filtered": 50.7, "entity_tracking": 21.34, "comps": 52.64, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "a50f988a565f5576d013095bc4571845e557769ce8424df1addb671d7e2b0d99", "precision": "fp32", "verification": "primary", "sources": [ "sf8aeec39c55c8b", "s83aae2815ca622" ] }, { "age": 80.0, "six": 49.09916666666667, "five": 50.88, "tasks": { "blimp_filtered": 69.66, "supplement_filtered": 60.13, "ewok_filtered": 50.62, "entity_tracking": 21.57, "comps": 52.42, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 60.0 }, "checkpoint": "443311723bdae0665924a41a14076076f294cb3ce48284c2c9bca288913f1b87", "precision": "fp32", "verification": "primary", "sources": [ "sbe780609714ad3", "s40e0c7a2461fa7" ] } ], "peak": { "five": { "score": 50.88, "ages": [ 80.0 ], "rank": 64 }, "six": { "score": 49.09916666666667, "ages": [ 80.0 ], "rank": 25 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 22, "id": "run/IPRLM320_FFN_ADDITIVE_BYPASS_SEED20260907_20261014", "original_name": "IPRLM320_FFN_ADDITIVE_BYPASS_SEED20260907_20261014", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After each40M training; no automatic extension.", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M training", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Both matchedseed parent320 at40M;40M allocates budget, does not establish final peak." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 79, "sources": [ "s45deda21d59fd0", "s9991b39ef31c03", "s8ad84476c3b04e" ], "points": [ { "age": 40.0, "six": 48.215833333333336, "five": 49.846, "tasks": { "blimp_filtered": 69.0, "supplement_filtered": 57.62, "ewok_filtered": 50.15, "entity_tracking": 20.02, "comps": 52.44, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 51.0 }, "checkpoint": "14755190bcee354fd64d0bcb5f3f53e343e6b7240994c8ccb7ac340b331aea4f", "precision": "fp32", "verification": "primary", "sources": [ "s0f3ee2bfc8f96d", "se97fd1dccb1346" ] }, { "age": 70.0, "six": 48.7775, "five": 51.0, "tasks": { "blimp_filtered": 70.58, "supplement_filtered": 58.66, "ewok_filtered": 50.92, "entity_tracking": 22.44, "comps": 52.4, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 53.0 }, "checkpoint": "64ee03a648803bf03fbab81c300225a0deb29382d98aa2a42356517f78075540", "precision": "fp32", "verification": "primary", "sources": [ "s15e0b4f51db919", "s3825c084a3ee66" ] }, { "age": 80.0, "six": 49.09583333333333, "five": 51.182, "tasks": { "blimp_filtered": 69.35, "supplement_filtered": 60.8, "ewok_filtered": 51.75, "entity_tracking": 21.62, "comps": 52.39, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 55.0 }, "checkpoint": "feba81c2ee1e90cc5a9dc7735572878e81f451af940ceb42350f52af24bf2cac", "precision": "fp32", "verification": "primary", "sources": [ "seea7686b316475", "s1b5488f98e794c" ] } ], "peak": { "five": { "score": 51.182, "ages": [ 80.0 ], "rank": 27 }, "six": { "score": 49.09583333333333, "ages": [ 80.0 ], "rank": 26 } }, "clear_points": [ { "number": 22, "age": 80, "seed": 20260907, "checkpoint_sha256": "feba81c2ee1e90cc5a9dc7735572878e81f451af940ceb42350f52af24bf2cac", "old_six": 49.09583333333333, "old_five": 51.182, "old_tasks": { "blimp_filtered": 69.35, "supplement_filtered": 60.8, "ewok_filtered": 51.75, "entity_tracking": 21.62, "comps": 52.39, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 55.0 }, "correct": 12315, "total": 17274, "accuracy": 71.29211531781868, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6950, "total": 8392, "ties": 0 }, "comps": { "correct": 3893, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 1, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 17, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 994, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 71, "total": 114, "ties": 0 }, "boolq": { "correct": 42, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 8, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 258, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 63.405034572999284, "checkpoint_rank": 5, "rank": 5 } ], "clear_peak": { "number": 22, "age": 80, "seed": 20260907, "checkpoint_sha256": "feba81c2ee1e90cc5a9dc7735572878e81f451af940ceb42350f52af24bf2cac", "old_six": 49.09583333333333, "old_five": 51.182, "old_tasks": { "blimp_filtered": 69.35, "supplement_filtered": 60.8, "ewok_filtered": 51.75, "entity_tracking": 21.62, "comps": 52.39, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 55.0 }, "correct": 12315, "total": 17274, "accuracy": 71.29211531781868, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6950, "total": 8392, "ties": 0 }, "comps": { "correct": 3893, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 1, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 17, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 994, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 71, "total": 114, "ties": 0 }, "boolq": { "correct": 42, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 8, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 258, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 63.405034572999284, "checkpoint_rank": 5, "rank": 5 }, "present": true }, { "number": 23, "id": "run/IPRLM384_LR_A_MAJOR95_OTHER100_SEED20260907_40M_20261020", "original_name": "IPRLM384_LR_A_MAJOR95_OTHER100_SEED20260907_40M_20261020", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要−5%・その他据置き", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005066666666666667, 0.0004433333333333333, 0.00037999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All15matchedseed40MLRconditions;reuse12controls;allgenres/pairedoutcomes/fullinventory" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 118, "sources": [ "sae69cd628b8209" ], "points": [ { "age": 40.0, "six": 48.376666666666665, "five": 50.334, "tasks": { "blimp_filtered": 68.15, "supplement_filtered": 60.81, "ewok_filtered": 50.59, "entity_tracking": 20.18, "comps": 51.94, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 50.0 }, "checkpoint": "2639172cbb513ac80ada9d659ffbb68251abb40943bc6f156c3c0d16e626fc71", "precision": "fp32", "verification": "primary", "sources": [ "s6c8248b694236d", "seea40757363ecd" ] }, { "age": 70.0, "six": 48.240833333333335, "five": 50.756, "tasks": { "blimp_filtered": 70.04, "supplement_filtered": 58.22, "ewok_filtered": 54.19, "entity_tracking": 19.58, "comps": 51.75, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 49.0 }, "checkpoint": "0766d1bfb6f05b29f845ee60ff100944af85a1778a377e98a725cdfcdd2cde32", "precision": "fp32", "verification": "primary", "sources": [ "s6c57328e392c07", "seea40757363ecd" ] }, { "age": 80.0, "six": 49.03916666666667, "five": 51.114, "tasks": { "blimp_filtered": 69.79, "supplement_filtered": 60.42, "ewok_filtered": 53.11, "entity_tracking": 19.86, "comps": 52.39, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 55.0 }, "checkpoint": "55659ba15037547c979dad6cce3261dea3bfccdfb1c498153a8c616f988b33d8", "precision": "fp32", "verification": "primary", "sources": [ "s701bcad4eaedcb", "seea40757363ecd" ] } ], "peak": { "five": { "score": 51.114, "ages": [ 80.0 ], "rank": 33 }, "six": { "score": 49.03916666666667, "ages": [ 80.0 ], "rank": 27 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 24, "id": "baby/B_READER_DEPTH9_WIDTH272_FAN_SCALED_NO_DISTANCE_100M", "original_name": "旧272・幅補正＋可変・距離補正なし", "title": "272幅・FFN960・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 272, "ffn": 960, "depth": 9, "parts": { "width": 272, "ffn": 960, "depth": 9, "attention": "8×34", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5294117647058822 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 272, "layers": 9, "hidden": 960, "attention_heads": 8, "head_dim": 34, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "history_norm": "separate", "seed": 20260907, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score60/70/80/90/100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 24, "sources": [ "s09c35c9a733cc2", "s7106887b67b746", "s7c126bd5a623fa", "s7e3e4643b1b306", "scf98b364e649e2", "se20384ea6bc3aa", "s25647f225e907f", "s946ca963c68dfc" ], "points": [ { "age": 60, "six": 47.73166666666667, "five": 49.454, "tasks": { "blimp_filtered": 67.84, "supplement_filtered": 56.9, "ewok_filtered": 51.7, "entity_tracking": 19.17, "comps": 51.66, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 53.0 }, "checkpoint": "5dc984e1d350feb83f9bdb02f84fc4d978026193ca912a450175574faffe0004", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "see5b9671b7a2f5", "s331a6a31840967", "sc61d97fcafc29d", "s49a1fc7254c81e", "sccec8896a2834f", "sd9a8d4b7b813d9", "s79ef61108d62bd", "s497cf6031ae80c", "sdd9eccd155c47c" ] }, { "age": 70, "six": 47.48916666666667, "five": 49.648, "tasks": { "blimp_filtered": 69.32, "supplement_filtered": 57.32, "ewok_filtered": 51.59, "entity_tracking": 17.95, "comps": 52.06, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 53.0 }, "checkpoint": "c5ce81dca3b7d97f2c6ec84993bb97f2566625d25ef5de7a4c4123053255fa83", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "see5b9671b7a2f5", "s331a6a31840967", "sb915ad8fc86c17", "sdad1f46df6f62a", "sccec8896a2834f", "sd9a8d4b7b813d9", "s79ef61108d62bd", "s497cf6031ae80c", "sdd9eccd155c47c" ] }, { "age": 80, "six": 49.035833333333336, "five": 50.616, "tasks": { "blimp_filtered": 68.55, "supplement_filtered": 61.04, "ewok_filtered": 51.48, "entity_tracking": 19.9, "comps": 52.11, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 58.0 }, "checkpoint": "de5902963a1d6880b00b4dd48b9a2c29a850e8932231e398f160599f536eff58", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "see5b9671b7a2f5", "s331a6a31840967", "s28e9a369a4cae3", "s7f2ff175402985", "sccec8896a2834f", "sd9a8d4b7b813d9", "s79ef61108d62bd", "s497cf6031ae80c", "sdd9eccd155c47c" ] }, { "age": 90, "six": 46.97, "five": 49.04, "tasks": { "blimp_filtered": 67.2, "supplement_filtered": 55.8, "ewok_filtered": 50.03, "entity_tracking": 19.97, "comps": 52.2, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 48.0 }, "checkpoint": "5d36f93adddbd4f14ad3c737df74dade307d0d96583792f0ef5e7a2395b59a66", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "see5b9671b7a2f5", "s331a6a31840967", "sb5eaa0d3f1d1d3", "s809015414bf1b0", "sccec8896a2834f", "sd9a8d4b7b813d9", "s79ef61108d62bd", "s497cf6031ae80c", "sdd9eccd155c47c" ] }, { "age": 100, "six": 47.755833333333335, "five": 50.186, "tasks": { "blimp_filtered": 69.53, "supplement_filtered": 57.22, "ewok_filtered": 52.26, "entity_tracking": 19.66, "comps": 52.26, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 45.0 }, "checkpoint": "bb95d72a9ca535453f979c2547a7bbafe823db0ea56177e9c5da93a8d54e1b38", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "see5b9671b7a2f5", "s331a6a31840967", "s52f12a73095033", "se4248f851cd28a", "sccec8896a2834f", "sd9a8d4b7b813d9", "s79ef61108d62bd", "s497cf6031ae80c", "sdd9eccd155c47c" ] } ], "peak": { "five": { "score": 50.616, "ages": [ 80 ], "rank": 95 }, "six": { "score": 49.035833333333336, "ages": [ 80 ], "rank": 29 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 25, "id": "run/IPRLM384_PREDICTOR504_MIXING432_SEED20260907_80M_20261023", "original_name": "IPRLM384_PREDICTOR504_MIXING432_SEED20260907_80M_20261023", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "予測7行列504／混合432（MoS6行列105版から2LR変更）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40/70/80M after80M training" ], "intermediate": "saved40/70M states evaluated after80M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000, 70000000, 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000, 70000000, 80000000 ], "timing": "after80M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All28LRconditions plusparent320/fullinventory;separate same-age and observedpeaks" }, "milestones_words": [ 40000000, 70000000, 80000000 ], "checkpoints_words": [ 40000000, 70000000, 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 139, "sources": [ "sbb261366b5f742", "sbb16090135352c" ], "points": [ { "age": 40.0, "six": 48.02583333333333, "five": 50.592, "tasks": { "blimp_filtered": 69.87, "supplement_filtered": 57.18, "ewok_filtered": 52.52, "entity_tracking": 20.57, "comps": 52.82, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 50.0 }, "checkpoint": "3ea94b584a570d2d41ede38f29df2346b8f96908c74481ff2fa1ad12657a7f48", "precision": "fp32", "verification": "primary", "sources": [ "sd27c13bbbf95b3", "sd6fc4a5e8ca93d" ] }, { "age": 70.0, "six": 49.03333333333333, "five": 51.398, "tasks": { "blimp_filtered": 70.49, "supplement_filtered": 59.95, "ewok_filtered": 52.49, "entity_tracking": 21.45, "comps": 52.61, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 55.0 }, "checkpoint": "4eada7d5abc2080383ab774120bab5c5c402f70923dc9feae415eb189e522f0d", "precision": "fp32", "verification": "primary", "sources": [ "seebf133f135206", "sd6fc4a5e8ca93d" ] }, { "age": 80.0, "six": 48.725, "five": 50.734, "tasks": { "blimp_filtered": 69.75, "supplement_filtered": 59.45, "ewok_filtered": 51.45, "entity_tracking": 20.62, "comps": 52.4, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 56.0 }, "checkpoint": "6a597c5c6dd3cf15661675d9b95d46fa1d6b3eacec3e7c1f6f8a1ddd6c8b1b8a", "precision": "fp32", "verification": "primary", "sources": [ "sd60393f1d0b4db", "sd6fc4a5e8ca93d" ] } ], "peak": { "five": { "score": 51.398, "ages": [ 70.0 ], "rank": 12 }, "six": { "score": 49.03333333333333, "ages": [ 70.0 ], "rank": 30 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 26, "id": "run/IPRLM320_PARENT_SEED137_20261015", "original_name": "IPRLM320_PARENT_SEED137_20261015", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at80M", "SyntaxGym31 and full PIQA1838", "frozen grounded2947" ], "intermediate": "No intermediate evaluation; train continuously to80M.", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 80000000 ], "timing": "after80M training", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Matched parent versus additive seed137 plus prior two matched pairs." }, "milestones_words": [ 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 77, "sources": [ "sf01bdc39b80224", "sea3d174fcc60fa", "s4ab85f37d005f8" ], "points": [ { "age": 40.0, "six": 47.155, "five": 50.036, "tasks": { "blimp_filtered": 68.9, "supplement_filtered": 58.09, "ewok_filtered": 51.16, "entity_tracking": 19.8, "comps": 52.23, "global_piqa_parallel": 16.5, "global_piqa_nonparallel": 49.0 }, "checkpoint": "d94ec34b68cc8a82c85ca87e26b3b29ae84ff811ed86346c239315cd1a050d29", "precision": "fp32", "verification": "primary", "sources": [ "s056520ba3174fe", "sad09599f2d0a93" ] }, { "age": 70.0, "six": 49.01083333333333, "five": 50.986, "tasks": { "blimp_filtered": 70.8, "supplement_filtered": 59.87, "ewok_filtered": 50.83, "entity_tracking": 20.47, "comps": 52.96, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 54.0 }, "checkpoint": "0900b070ac6f41bdd03e92c7704969e9ed082565591e65ac1beb886a8fbd4ede", "precision": "fp32", "verification": "primary", "sources": [ "s0497e33afa03a0", "s7e10330e6c3b8e" ] }, { "age": 80.0, "six": 48.111666666666665, "five": 51.092, "tasks": { "blimp_filtered": 70.1, "supplement_filtered": 60.56, "ewok_filtered": 53.3, "entity_tracking": 19.08, "comps": 52.42, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 47.0 }, "checkpoint": "e71398961a0f2d74f8bb418b936cc5370fbce2e8ccc93e223fd1b1448a3f7d32", "precision": "fp32", "verification": "primary", "sources": [ "s8d92dc8a537e60", "s7ded66e1bb5b24" ] } ], "peak": { "five": { "score": 51.092, "ages": [ 80.0 ], "rank": 35 }, "six": { "score": 49.01083333333333, "ages": [ 70.0 ], "rank": 33 } }, "clear_points": [ { "number": 26, "age": 70, "seed": 137, "checkpoint_sha256": "0900b070ac6f41bdd03e92c7704969e9ed082565591e65ac1beb886a8fbd4ede", "old_six": 49.01083333333333, "old_five": 50.986, "old_tasks": { "blimp_filtered": 70.8, "supplement_filtered": 59.87, "ewok_filtered": 50.83, "entity_tracking": 20.47, "comps": 52.96, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 54.0 }, "correct": 12256, "total": 17274, "accuracy": 70.9505615375709, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6930, "total": 8392, "ties": 0 }, "comps": { "correct": 3917, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 3, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 18, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 996, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 70, "total": 114, "ties": 0 }, "boolq": { "correct": 21, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 62.05388833300094, "checkpoint_rank": 8 }, { "number": 26, "age": 80, "seed": 137, "checkpoint_sha256": "e71398961a0f2d74f8bb418b936cc5370fbce2e8ccc93e223fd1b1448a3f7d32", "old_six": 48.111666666666665, "old_five": 51.092, "old_tasks": { "blimp_filtered": 70.1, "supplement_filtered": 60.56, "ewok_filtered": 53.3, "entity_tracking": 19.08, "comps": 52.42, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 47.0 }, "correct": 12373, "total": 17274, "accuracy": 71.62788005094362, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6856, "total": 8392, "ties": 0 }, "comps": { "correct": 3909, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 20, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 998, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 72, "total": 114, "ties": 0 }, "boolq": { "correct": 13, "total": 53, "ties": 0 }, "mnli": { "correct": 29, "total": 203, "ties": 0 }, "mrpc": { "correct": 7, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 410, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 62.007026354003756, "checkpoint_rank": 4, "rank": 4 } ], "clear_peak": { "number": 26, "age": 80, "seed": 137, "checkpoint_sha256": "e71398961a0f2d74f8bb418b936cc5370fbce2e8ccc93e223fd1b1448a3f7d32", "old_six": 48.111666666666665, "old_five": 51.092, "old_tasks": { "blimp_filtered": 70.1, "supplement_filtered": 60.56, "ewok_filtered": 53.3, "entity_tracking": 19.08, "comps": 52.42, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 47.0 }, "correct": 12373, "total": 17274, "accuracy": 71.62788005094362, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6856, "total": 8392, "ties": 0 }, "comps": { "correct": 3909, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 20, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 998, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 72, "total": 114, "ties": 0 }, "boolq": { "correct": 13, "total": 53, "ties": 0 }, "mnli": { "correct": 29, "total": 203, "ties": 0 }, "mrpc": { "correct": 7, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 410, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 62.007026354003756, "checkpoint_rank": 4, "rank": 4 }, "present": true }, { "number": 27, "id": "run/IPRLM384_JOINT_ATTENTION_READER_LR_SEED20260907_80M_20261022", "original_name": "IPRLM384_JOINT_ATTENTION_READER_LR_SEED20260907_80M_20261022", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "V/out504・Router552・Reader780/gate440", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40/70/80M after80M training" ], "intermediate": "saved40/70M states evaluated after80M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000, 70000000, 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000, 70000000, 80000000 ], "timing": "after80M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All25LRconditions and fullmodelinventory;matchedRouter125parent" }, "milestones_words": [ 40000000, 70000000, 80000000 ], "checkpoints_words": [ 40000000, 70000000, 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 109, "sources": [ "s21753b625caa7c", "saadfdc6fd7781a" ], "points": [ { "age": 40.0, "six": 47.38, "five": 50.314, "tasks": { "blimp_filtered": 69.63, "supplement_filtered": 58.89, "ewok_filtered": 51.99, "entity_tracking": 18.94, "comps": 52.12, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 46.0 }, "checkpoint": "3b22f2c3330eeed3131afabfb9f6f04bf9361d9f0173f50b2b6cd6f31aa1b021", "precision": "fp32", "verification": "primary", "sources": [ "s5d755e7304bd1a", "sc60026e88ac314" ] }, { "age": 70.0, "six": 49.0075, "five": 51.57, "tasks": { "blimp_filtered": 70.29, "supplement_filtered": 61.82, "ewok_filtered": 52.9, "entity_tracking": 20.35, "comps": 52.49, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 52.0 }, "checkpoint": "af2dc63af532044d9193a4cdbb29335df0fe3dfade22ef21135a43e74be503e9", "precision": "fp32", "verification": "primary", "sources": [ "se9db67c365e86a", "sc60026e88ac314" ] }, { "age": 80.0, "six": 48.80916666666667, "five": 51.044, "tasks": { "blimp_filtered": 69.68, "supplement_filtered": 61.8, "ewok_filtered": 51.28, "entity_tracking": 20.03, "comps": 52.43, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 51.0 }, "checkpoint": "eb3f4d90cbdd23f9396e422f83f1f0322f9cdfeb023c1be5b0caf6f65352cddf", "precision": "fp32", "verification": "primary", "sources": [ "scb112ddc080271", "sc60026e88ac314" ] } ], "peak": { "five": { "score": 51.57, "ages": [ 70.0 ], "rank": 7 }, "six": { "score": 49.0075, "ages": [ 70.0 ], "rank": 34 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 28, "id": "baby/vistarc_q3_mos3_last2_width272_ffn960_attn256_20261007", "original_name": "272・MoS3・2層混合・Attention256", "title": "272幅・FFN960・9層", "topology": "MoS3・共通2層混合", "recipe": "主要0.0007529／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 272, "ffn": 960, "depth": 9, "parts": { "width": 272, "ffn": 960, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 8, 9 ] ], "ffn_ratio": 3.5294117647058822 }, "main_lr": [ 0.0007529411764705883, 0.0006588235294117646, 0.0005647058823529411 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 272, "layers": 9, "hidden": 960, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last2", "attention_received_gate": true, "output_components": 3, "output_state_layers": 2, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3_LAST2 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 65536, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.315, "comparison_scope": "last2_256 parent plus last3 and last4 reference context" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 50, "sources": [ "s2d667250923fc8", "sebcb6e92988178", "s556b5ecffe38a6", "s4997da61cf83e8", "s7962c241145090", "sb6a2f463211859" ], "points": [ { "age": 40.0, "six": 47.958333333333336, "five": 50.326, "tasks": { "blimp_filtered": 69.7, "supplement_filtered": 57.87, "ewok_filtered": 52.26, "entity_tracking": 20.14, "comps": 51.66, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 47.0 }, "checkpoint": "a60f90f04cb45181b754a21b728f02feffd597de8641749b108a67a3fb5f401d", "precision": "fp32", "verification": "primary", "sources": [ "s7a692edec4c712", "s0043626eb18246" ] }, { "age": 70, "six": 48.6925, "five": 50.804, "tasks": { "blimp_filtered": 70.56, "supplement_filtered": 59.69, "ewok_filtered": 51.19, "entity_tracking": 20.74, "comps": 51.84, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 52.0 }, "checkpoint": "f16f5403a38c5b154ab57bdf03ba7a50a0d0d893ad0e1dc55ae4668e7c81b9b1", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4a62d404b50561", "se7000fa2361b32", "s60b2baa9d04406", "s8b5bc4d4304436", "s009cbd93cc5b3b", "se5c5ef9e7604d1" ] }, { "age": 80, "six": 48.998333333333335, "five": 50.874, "tasks": { "blimp_filtered": 68.97, "supplement_filtered": 61.72, "ewok_filtered": 50.6, "entity_tracking": 21.19, "comps": 51.89, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 54.0 }, "checkpoint": "77d6512bcb6c6bf0f00714bc11ac29c378868acd985cd09ab74c7fc886d524b6", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s7be3509ef472c9", "sd70281faf3ef0f", "s60b2baa9d04406", "s8b5bc4d4304436", "s009cbd93cc5b3b", "se5c5ef9e7604d1" ] }, { "age": 90, "six": 48.1475, "five": 50.45, "tasks": { "blimp_filtered": 70.13, "supplement_filtered": 57.35, "ewok_filtered": 51.49, "entity_tracking": 21.4, "comps": 51.88, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 49.0 }, "checkpoint": "382b5c6f7c23be89dd93475dd6cdb7cae7bba1324f0a8dbe7246c0094f52ada6", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sf7b26987487ac2", "sac5efa75189eca", "s60b2baa9d04406", "s8b5bc4d4304436", "s009cbd93cc5b3b", "se5c5ef9e7604d1" ] }, { "age": 100, "six": 48.585, "five": 50.578, "tasks": { "blimp_filtered": 70.16, "supplement_filtered": 59.5, "ewok_filtered": 50.95, "entity_tracking": 20.07, "comps": 52.21, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 52.0 }, "checkpoint": "441c5116f5f9596f99f59ef7df1c73efd9234bef98f5b3e19f255ab1164e21ef", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s460f29091e5318", "s2fda18dd2c15d0", "s60b2baa9d04406", "s8b5bc4d4304436", "s009cbd93cc5b3b", "se5c5ef9e7604d1" ] } ], "peak": { "five": { "score": 50.874, "ages": [ 80 ], "rank": 65 }, "six": { "score": 48.998333333333335, "ages": [ 80 ], "rank": 35 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 29, "id": "run/IPRLM384_NORM820_READER_DOWN800_GATE420_SEED20260907_40M_20261022", "original_name": "IPRLM384_NORM820_READER_DOWN800_GATE420_SEED20260907_40M_20261022", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "Norm820＋Reader800／ゲート420", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All23same-seed40Mconditions;all7tasksandpairedoutcomes;fullknowninventory;deltasagainstoriginalNorm820" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 136, "sources": [ "s823a6845d3017a", "s90429899c71135" ], "points": [ { "age": 40.0, "six": 47.94833333333333, "five": 50.596, "tasks": { "blimp_filtered": 69.05, "supplement_filtered": 59.77, "ewok_filtered": 51.83, "entity_tracking": 20.08, "comps": 52.25, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 50.0 }, "checkpoint": "68768a860de38a684d5035d8284dfb65efa8031e934fc97f1ac43cd5aa010420", "precision": "fp32", "verification": "primary", "sources": [ "sae961e7f59d245", "s1f1540b9535ae1" ] }, { "age": 70.0, "six": 48.9875, "five": 51.164, "tasks": { "blimp_filtered": 70.26, "supplement_filtered": 62.23, "ewok_filtered": 51.04, "entity_tracking": 19.68, "comps": 52.61, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 50.0 }, "checkpoint": "2d341821530ffda8049d3f5eff83a78dfceac6e200e5a3f68a61282f3cfe1e58", "precision": "fp32", "verification": "primary", "sources": [ "s22c62debb43412", "s1f1540b9535ae1" ] }, { "age": 80.0, "six": 48.925, "five": 51.462, "tasks": { "blimp_filtered": 69.7, "supplement_filtered": 63.05, "ewok_filtered": 52.63, "entity_tracking": 19.47, "comps": 52.46, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 55.0 }, "checkpoint": "bf8314d12041d86e57dd2d7f0a85a42ce23bf95c004a1e2fb91585d168910ff4", "precision": "fp32", "verification": "primary", "sources": [ "s646a03d83f9771", "s1f1540b9535ae1" ] } ], "peak": { "five": { "score": 51.462, "ages": [ 80.0 ], "rank": 10 }, "six": { "score": 48.9875, "ages": [ 70.0 ], "rank": 36 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 30, "id": "run/IPRLM384_MAJOR85_OTHER105_SEED20260907_40M_20261024", "original_name": "IPRLM384_MAJOR85_OTHER105_SEED20260907_40M_20261024", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要−15％・その他＋5％", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00045333333333333337, 0.00039666666666666664, 0.00033999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All known same-age conditions and complete known inventory;B/unchanged384/parent320 matched40M references;40M is a budget decision,not peak determination" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 121, "sources": [ "sd8c5630f289215" ], "points": [ { "age": 40.0, "six": 48.315, "five": 50.442, "tasks": { "blimp_filtered": 69.34, "supplement_filtered": 57.68, "ewok_filtered": 52.89, "entity_tracking": 19.83, "comps": 52.47, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 54.0 }, "checkpoint": "e9a8f69edbe128a84ffa4002bbd3cb413602a82cd51e85de0dd5850984deb4a0", "precision": "fp32", "verification": "primary", "sources": [ "sc1153917461f92", "s6a7336d3bcb015" ] }, { "age": 70.0, "six": 48.876666666666665, "five": 51.022, "tasks": { "blimp_filtered": 69.86, "supplement_filtered": 59.27, "ewok_filtered": 52.72, "entity_tracking": 21.25, "comps": 52.01, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 53.0 }, "checkpoint": "719382fa2beba35b68f915b2d83316ee01466e6b69c5171e0aef1c201147e36b", "precision": "fp32", "verification": "primary", "sources": [ "s9b86f1b6e72cd5", "s6a7336d3bcb015" ] }, { "age": 80.0, "six": 48.96333333333333, "five": 51.114, "tasks": { "blimp_filtered": 68.57, "supplement_filtered": 61.59, "ewok_filtered": 52.71, "entity_tracking": 20.46, "comps": 52.24, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 57.0 }, "checkpoint": "f2a972e6fb344a1d6d629ec691f02172b1bf10fdc7dec968de9398b733dccdeb", "precision": "fp32", "verification": "primary", "sources": [ "sc41cbeef51ac33", "s6a7336d3bcb015" ] } ], "peak": { "five": { "score": 51.114, "ages": [ 80.0 ], "rank": 33 }, "six": { "score": 48.96333333333333, "ages": [ 80.0 ], "rank": 38 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 31, "id": "run/IPRLM384_FFN1408_MAJOR_LR533_SEED20260907_20261019", "original_name": "IPRLM384_FFN1408_MAJOR_LR533_SEED20260907_20261019", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "基準384：主要533・その他基準", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005333333333333334, 0.00046666666666666666, 0.00039999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at80M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 80000000 ], "timing": "serial after80M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Full known inventory plus matched-age parent320; observed peaks retained separately" }, "milestones_words": [ 80000000 ], "checkpoints_words": [ 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 105, "sources": [ "s1b9b822b32a4c4", "s64611ca4a8fce7" ], "points": [ { "age": 40.0, "six": 48.285833333333336, "five": 50.316, "tasks": { "blimp_filtered": 70.0, "supplement_filtered": 57.38, "ewok_filtered": 51.46, "entity_tracking": 20.37, "comps": 52.37, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 52.0 }, "checkpoint": "526309b3a8ccc740542669b919ee8034ea8c543b0a279fc82a30b147d4491d43", "precision": "fp32", "verification": "primary", "sources": [ "s9c8077aebeff13" ] }, { "age": 70.0, "six": 48.854166666666664, "five": 50.618, "tasks": { "blimp_filtered": 69.57, "supplement_filtered": 58.76, "ewok_filtered": 50.92, "entity_tracking": 21.62, "comps": 52.22, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 49.0 }, "checkpoint": "59545b82937cc880e9ecc4f7279557dadb88a82bc8d05b64b34c8759633e6867", "precision": "fp32", "verification": "primary", "sources": [ "s0906fc50f8932f" ] }, { "age": 80.0, "six": 48.96, "five": 50.634, "tasks": { "blimp_filtered": 69.46, "supplement_filtered": 60.44, "ewok_filtered": 49.57, "entity_tracking": 21.28, "comps": 52.42, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 54.0 }, "checkpoint": "5a43de54b04b42cc5d25c0f1a71360f711bcc105e038fcb0077ae1a5443de649", "precision": "fp32", "verification": "primary", "sources": [ "sc33d5af2186f23", "s7408bcd2016ad2" ] } ], "peak": { "five": { "score": 50.634, "ages": [ 80.0 ], "rank": 90 }, "six": { "score": 48.96, "ages": [ 80.0 ], "rank": 39 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 32, "id": "baby/vistarc_q3_mos3_last1_width512_ffn1792_heads16_20261008", "original_name": "512・MoS3・1層参照", "title": "512幅・FFN1792・9層", "topology": "MoS3", "recipe": "主要0.0004／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 512, "ffn": 1792, "depth": 9, "parts": { "width": 512, "ffn": 1792, "depth": 9, "attention": "16×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "最終層", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.0004, 0.00035, 0.0003 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 512, "layers": 9, "hidden": 1792, "attention_heads": 16, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 512, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last1", "attention_received_gate": true, "output_components": 3, "output_state_layers": 1, "reference_attention_width": 256, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 exact recurrent scoring;512 runtime acceptance pending", "runtime_selection": { "graph_chunk": 4, "max_batch": 512, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.328, "same_family_peak_six_axis": 49.315, "comparison_scope": "last1_384 parent plus last2_384 and last2_256", "parent_peak_six_axis": 49.37916666666667 }, "milestones_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ], "status": "AUTHORIZED_PENDING_SPEED_ACCEPTANCE" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 48, "sources": [ "sf3619173a9eaf5", "sc2fd32f80e3d2c", "s8fedd9b8f9bc7d", "sbd1b86dccea413", "se5ccda33981d26" ], "points": [ { "age": 50, "six": 47.6025, "five": 50.678, "tasks": { "blimp_filtered": 70.18, "supplement_filtered": 58.72, "ewok_filtered": 52.82, "entity_tracking": 19.38, "comps": 52.29, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 46.0 }, "checkpoint": "d4cdae7bffa850f816f7eac21bb7d146c584ec616a7a6f77d977ece8372bca13", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s36f8e2dc1f0a1e", "s1c87bb6c5a397c", "s5bfaee329a124e", "sc73cddbe4fbe6e" ] }, { "age": 60, "six": 47.6075, "five": 49.908, "tasks": { "blimp_filtered": 67.88, "supplement_filtered": 58.6, "ewok_filtered": 51.78, "entity_tracking": 19.32, "comps": 51.96, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 46.0 }, "checkpoint": "e202f746063d9aa2c4a1b5078c974d7e12e076ad450abe9176f71ae3d63e2512", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s187cf237736681", "s1c87bb6c5a397c", "s5bfaee329a124e", "sc73cddbe4fbe6e" ] }, { "age": 70, "six": 48.55, "five": 51.23, "tasks": { "blimp_filtered": 69.61, "supplement_filtered": 61.03, "ewok_filtered": 53.24, "entity_tracking": 20.19, "comps": 52.08, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 47.0 }, "checkpoint": "daf65cdd2778218202b74bf8d68dcfabc3c19f33de4da2c20453f783a57c6038", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s7e3935da1f4cbd", "s1c87bb6c5a397c", "s5bfaee329a124e", "sc73cddbe4fbe6e" ] }, { "age": 80, "six": 48.946666666666665, "five": 51.2, "tasks": { "blimp_filtered": 68.93, "supplement_filtered": 61.7, "ewok_filtered": 52.14, "entity_tracking": 21.03, "comps": 52.2, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 54.0 }, "checkpoint": "9f11b2a377ce426003632cb55fd04d6c9487dea8daa08982f09f07bc1a8bd8f3", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s56511db0dca097", "s1c87bb6c5a397c", "s5bfaee329a124e", "sc73cddbe4fbe6e" ] }, { "age": 90, "six": 47.1025, "five": 50.084, "tasks": { "blimp_filtered": 68.02, "supplement_filtered": 57.37, "ewok_filtered": 52.45, "entity_tracking": 20.36, "comps": 52.22, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 44.0 }, "checkpoint": "bef914e982a4b7ca770b12c8f52457a481254a2dbbbc05a194b382230b401d9c", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sa1bafec8306b6a", "s1c87bb6c5a397c", "s5bfaee329a124e", "sc73cddbe4fbe6e" ] }, { "age": 100, "six": 47.494166666666665, "five": 49.76, "tasks": { "blimp_filtered": 68.08, "supplement_filtered": 58.66, "ewok_filtered": 50.21, "entity_tracking": 19.56, "comps": 52.29, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 50.0 }, "checkpoint": "c177e7349525706c53d2790869bdc476977915f85108d70217b88d428472e83b", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc4381e8533565d", "s1c87bb6c5a397c", "s5bfaee329a124e", "sc73cddbe4fbe6e" ] } ], "peak": { "five": { "score": 51.23, "ages": [ 70 ], "rank": 24 }, "six": { "score": 48.946666666666665, "ages": [ 80 ], "rank": 40 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 33, "id": "run/VISTA336_CANDIDATE_READ_RECEIVED_GATE_MOS3_SEED20260907_20261018", "original_name": "VISTA336_CANDIDATE_READ_RECEIVED_GATE_MOS3_SEED20260907_20261018", "title": "336幅・FFN1232・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": "未確認", "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTA_CANDIDATE_READ_RECEIVED_GATE_MOS3", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "mos3", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": false, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "40M only after firstseed training; no remainingseed auto-launch", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 three-branch MoS3; same candidate feedback and scores; no training fused mixture during evaluation.", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M received-gate parent, FFNadd, Reader48, candidate-only and old parent336; no current benchmark result yet." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 161, "sources": [ "sd58880aae071e9", "sd7f627a6e84c24", "sa672af93df3a31" ], "points": [ { "age": 40.0, "six": 47.74333333333333, "five": 50.75, "tasks": { "blimp_filtered": 68.85, "supplement_filtered": 57.39, "ewok_filtered": 53.06, "entity_tracking": 22.1, "comps": 52.35, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 46.0 }, "checkpoint": "1098c4b74e1adad3aeb5b424709c0e021eb69fa05393d2cc6dbc1fc826aa0a99", "precision": "fp32", "verification": "primary", "sources": [ "s4ceb65a4affe2b", "se90a9b5e2e7871" ] }, { "age": 70.0, "six": 48.939166666666665, "five": 51.1, "tasks": { "blimp_filtered": 70.42, "supplement_filtered": 58.86, "ewok_filtered": 52.16, "entity_tracking": 21.6, "comps": 52.46, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 52.0 }, "checkpoint": "6c46c77ae51e4bfb2c1c414b3f20eb089f3547102b9d042a1be2b4e8e28894e2", "precision": "fp32", "verification": "primary", "sources": [ "sb5d4d6889806fd", "sa1ea9f5899c81c" ] }, { "age": 80.0, "six": 48.935833333333335, "five": 51.19, "tasks": { "blimp_filtered": 68.66, "supplement_filtered": 61.21, "ewok_filtered": 52.46, "entity_tracking": 20.84, "comps": 52.78, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 53.0 }, "checkpoint": "67c8499015c66307b6137de6ef81664ddfade834a601e91690e3ab5853644c97", "precision": "fp32", "verification": "primary", "sources": [ "sca6229c92463bd", "sa7d78f688525c4" ] }, { "age": 90.0, "six": 48.63583333333333, "five": 50.336, "tasks": { "blimp_filtered": 69.23, "supplement_filtered": 56.79, "ewok_filtered": 52.69, "entity_tracking": 20.49, "comps": 52.48, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 56.0 }, "checkpoint": "88eb5aff8c446eba76f961384ded7d3d3a00a437891867a714e63c104504ea4b", "precision": "fp32", "verification": "primary", "sources": [ "s9815eb47d9017a", "sd6c1617c9f028c" ] }, { "age": 100.0, "six": 48.56333333333333, "five": 50.746, "tasks": { "blimp_filtered": 68.64, "supplement_filtered": 58.62, "ewok_filtered": 52.25, "entity_tracking": 21.72, "comps": 52.5, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 52.0 }, "checkpoint": "1067f9e1d9e74791cb233e10a11fc3f2d4b5891ce86a40a1ec8d4561d6658cd2", "precision": "fp32", "verification": "primary", "sources": [ "s63847a38f0e9ea", "s8932c8de5588f7" ] } ], "peak": { "five": { "score": 51.19, "ages": [ 80.0 ], "rank": 25 }, "six": { "score": 48.939166666666665, "ages": [ 70.0 ], "rank": 42 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 34, "id": "run/IPRLM_9L320_PREDICTION_AWARE_MIXTURE_GATE_20261011", "original_name": "IPRLM_9L320_PREDICTION_AWARE_MIXTURE_GATE_20261011", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "prediction_residual_v1", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_init_only", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123_prediction_gate", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "initialization_layers": 9, "output_mixture_gate": "prediction_residual_v1", "mixture_gate_hidden": 320, "mixture_gate_statistics": "normalized_entropy_js_v1", "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes70/80/90/100M", "SyntaxGym31+PIQA1838 at70/80/90/100M", "bAbI1947+ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 TF32off exact prefix graph; actual9-layer320 prediction gate speed, score parity and all capacity buckets verified", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Matched ages and observed peaks separately, including ordinary320 shared-last2 and all selected historical models" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ], "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true, "compile_prediction_statistics": true } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 143, "sources": [ "s640edf6f4628e0", "sb118803f312ece", "s3cf2c287bca401", "sd36fc5df11324f", "s4aa2a5afcf31cb", "s88e0a25eba109a" ], "points": [ { "age": 70.0, "six": 47.82083333333333, "five": 50.558, "tasks": { "blimp_filtered": 68.56, "supplement_filtered": 58.72, "ewok_filtered": 52.21, "entity_tracking": 21.15, "comps": 52.15, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 44.0 }, "checkpoint": "b7eecf4f95f904f77f8b371e2243568f04d634b12d388258be86fdf1eb13c405", "precision": "fp32", "verification": "primary", "sources": [ "sa5c58ab4ba602c", "sb49c34e60ac130", "sa9112d922a9096", "s967f032239c65b", "s648f69712d5799", "s67714866eedeb2" ] }, { "age": 80.0, "six": 48.93, "five": 51.0, "tasks": { "blimp_filtered": 68.04, "supplement_filtered": 60.59, "ewok_filtered": 52.4, "entity_tracking": 21.63, "comps": 52.34, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 49.0 }, "checkpoint": "e1c9b15367e792bf7c204f9080987c5b4788bc4dd866ca6428cbebbc18085247", "precision": "fp32", "verification": "primary", "sources": [ "sa2dc293a52c439", "sa07e5bd9e1fe17", "sa9112d922a9096", "s967f032239c65b", "s648f69712d5799", "s67714866eedeb2" ] }, { "age": 90.0, "six": 47.94166666666667, "five": 50.1, "tasks": { "blimp_filtered": 68.07, "supplement_filtered": 56.26, "ewok_filtered": 52.77, "entity_tracking": 21.27, "comps": 52.13, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "797d4d25d39146926641b7fa499fb9b3f8ebee26585de4d04c42862e8a664663", "precision": "fp32", "verification": "primary", "sources": [ "s3d1b6148042df9", "sbaba7ed4f15d94", "sa9112d922a9096", "s967f032239c65b", "s648f69712d5799", "s67714866eedeb2" ] }, { "age": 100.0, "six": 47.40833333333333, "five": 50.366, "tasks": { "blimp_filtered": 68.15, "supplement_filtered": 58.69, "ewok_filtered": 51.88, "entity_tracking": 20.89, "comps": 52.22, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 40.0 }, "checkpoint": "34f53fa067456d8b7d38b436589e4c9f2c4035cce05c02209888597240da79c6", "precision": "fp32", "verification": "primary", "sources": [ "sb8599bca64c001", "sa4b90ee2ed2ab9", "sa9112d922a9096", "s967f032239c65b", "s648f69712d5799", "s67714866eedeb2" ] } ], "peak": { "five": { "score": 51.0, "ages": [ 80.0 ], "rank": 46 }, "six": { "score": 48.93, "ages": [ 80.0 ], "rank": 43 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 35, "id": "run/IPRLM336_NO_READER_NO_FFN_ADD_SEED20260907_20261018", "original_name": "IPRLM336_NO_READER_NO_FFN_ADD_SEED20260907_20261018", "title": "336幅・FFN1232・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "帰還なし", "reader": "なし", "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTA_RECEIVED_MOS3_NO_FEEDBACK", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 1, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": false, "candidate_reader_enabled": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at70/80/90/100M after100M training" ], "intermediate": "No benchmarks during training; retain10M full states.", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact one-pass causal FP32 scoring, same questions, ranker, masks and length handling; no recurrent prediction feedback.", "runtime_selection": { "parallel_causal": true, "precision": "fp32", "tf32": false, "token_chunk": 512, "max_batch": 32, "kv_token_budget": 4096, "merge_blocks": 8 }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed age-matched old parent336 where available and new candidate-read MoS3 arm; report observed peaks and coverage separately." }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated causal attention; original adjacent lines inside groups share context; no prediction feedback", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 93, "sources": [ "s5022b64b2f9975", "sd7ab2acfcbe978", "sfa2774340d2cae" ], "points": [ { "age": 40.0, "six": 48.450833333333335, "five": 50.02, "tasks": { "blimp_filtered": 68.63, "supplement_filtered": 58.09, "ewok_filtered": 50.77, "entity_tracking": 20.25, "comps": 52.36, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 55.0 }, "checkpoint": "fda9c9ab08f2b34d754b1b65a327701265bdb39379745a2a7b1af5e84197a1eb", "precision": "fp32", "verification": "primary", "sources": [ "sf4b0340205cf70", "sa62c16ebbc210a" ] }, { "age": 70.0, "six": 48.44833333333333, "five": 50.714, "tasks": { "blimp_filtered": 69.41, "supplement_filtered": 58.89, "ewok_filtered": 51.98, "entity_tracking": 21.36, "comps": 51.93, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 49.0 }, "checkpoint": "5ee375e138da9f9b7f2eca3f2285b61bc88dc1102aa430b25e9ca4086f57e823", "precision": "fp32", "verification": "primary", "sources": [ "s97ff36e966c9df", "s86a9bedb9c7441" ] }, { "age": 80.0, "six": 48.924166666666665, "five": 51.082, "tasks": { "blimp_filtered": 68.78, "supplement_filtered": 61.66, "ewok_filtered": 51.39, "entity_tracking": 21.43, "comps": 52.15, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 52.0 }, "checkpoint": "794a8ad99c850ccb214fea37656fe5bf05cbe795d93544a7a0a8adf4e5c40417", "precision": "fp32", "verification": "primary", "sources": [ "s1ee4f5fc4caabf", "s29a67c201b6eed" ] }, { "age": 90.0, "six": 47.9675, "five": 50.54, "tasks": { "blimp_filtered": 69.01, "supplement_filtered": 58.81, "ewok_filtered": 51.59, "entity_tracking": 21.01, "comps": 52.28, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 44.0 }, "checkpoint": "9c0e7ddb2c3c8b54525cbe41bf6326d289cdb463c6693a1d2ee1f6415fbd13a4", "precision": "fp32", "verification": "primary", "sources": [ "s6f425ce200fb0b", "s47d83ee8831939" ] }, { "age": 100.0, "six": 48.75666666666667, "five": 50.792, "tasks": { "blimp_filtered": 68.98, "supplement_filtered": 58.66, "ewok_filtered": 51.69, "entity_tracking": 21.77, "comps": 52.86, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 49.0 }, "checkpoint": "48b6650c8e8b567aa8679dbfc15b3f154ab3e5b9664d2c0ef1d8020b254181d5", "precision": "fp32", "verification": "primary", "sources": [ "s83f498250b3417", "sdd3eec4c05f258" ] } ], "peak": { "five": { "score": 51.082, "ages": [ 80.0 ], "rank": 38 }, "six": { "score": 48.924166666666665, "ages": [ 80.0 ], "rank": 44 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 36, "id": "run/IPRLM320_FFN_CONTEXTUAL_ADDITION_SEED20260907_20261019", "original_name": "IPRLM320_FFN_CONTEXTUAL_ADDITION_SEED20260907_20261019", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true, "ffn_contextual_bypass": true, "ffn_context_scale": 0.003125 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at80M" ], "intermediate": "No intermediate benchmarks requested.", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph; bounded live decoder cache; batch sizing measured before training", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 80000000 ], "timing": "after80M training", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed80M parent320 and unconditional addition plus all-model inventory; one seed does not establish robustness." }, "milestones_words": [ 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 80, "sources": [ "s9616198b01a4aa", "s68ace6af45def2" ], "points": [ { "age": 80.0, "six": 48.880833333333335, "five": 50.824, "tasks": { "blimp_filtered": 69.02, "supplement_filtered": 60.67, "ewok_filtered": 52.35, "entity_tracking": 19.58, "comps": 52.5, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 56.0 }, "checkpoint": null, "precision": null, "verification": "primary", "sources": [ "s317bdcd68b4bcf" ] }, { "age": 80.0, "six": 48.880833333333335, "five": 50.824, "tasks": { "blimp_filtered": 69.02, "supplement_filtered": 60.67, "ewok_filtered": 52.35, "entity_tracking": 19.58, "comps": 52.5, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 56.0 }, "checkpoint": "b85ea32bd6f8667cbf160c995b94785041523973bbd7ca93a2aab13b1de01471", "precision": "fp32", "verification": "primary", "sources": [ "sc439afadb1896f" ] } ], "peak": { "five": { "score": 50.824, "ages": [ 80.0, 80.0 ], "rank": 73 }, "six": { "score": 48.880833333333335, "ages": [ 80.0, 80.0 ], "rank": 45 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 37, "id": "baby/vistarc_q3_mos9_layerwise_width384_ffn1344_heads12_20261009", "original_name": "384・MoS9・全9層分離", "title": "384幅・FFN1344・9層", "topology": "MoS9・各層分離", "recipe": "主要0.0005333／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1344, "depth": 9, "parts": { "width": 384, "ffn": 1344, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS9", "reference": "1層 / 2層 / 3層 / 4層 / 5層 / 6層 / 7層 / 8層 / 9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 9, "reference_mode": "各層を個別", "reference_layers": [ [ 1 ], [ 2 ], [ 3 ], [ 4 ], [ 5 ], [ 6 ], [ 7 ], [ 8 ], [ 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.0005333333333333334, 0.00046666666666666666, 0.00039999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1344, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos9_layerwise", "attention_received_gate": true, "output_components": 9, "output_state_layers": 9, "reference_attention_width": 256, "output_branch_layers": [ [ 1 ], [ 2 ], [ 3 ], [ 4 ], [ 5 ], [ 6 ], [ 7 ], [ 8 ], [ 9 ] ], "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph; native9-branch normalization, score and cache verified before launch", "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Current384 MoS9 versus320 branch123,384 common-last2 and full historical references; matched ages and observed peaks separately" }, "milestones_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 61, "sources": [ "se61287a95db33c", "s146e0883ba3d8d", "s0b1698817d99bc", "s87335f6b2c6c98", "sd8e7d3319a5f19", "s9c0f78a99c5fea" ], "points": [ { "age": 50, "six": 48.2225, "five": 51.128, "tasks": { "blimp_filtered": 69.04, "supplement_filtered": 61.36, "ewok_filtered": 50.76, "entity_tracking": 21.75, "comps": 52.73, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "92ba057ca276d50badd460a4abb2bdcd7f887ce7b0ef5c3700710fbccf1b77ca", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1537003e02d8ca", "s7eabfe190bde65", "s51972e8ac79cc7", "sbc8e36698eb403", "s5d922eccc266ec", "scf0385959efbd2" ] }, { "age": 60, "six": 47.4475, "five": 50.404, "tasks": { "blimp_filtered": 67.74, "supplement_filtered": 60.14, "ewok_filtered": 52.18, "entity_tracking": 19.98, "comps": 51.98, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 43.0 }, "checkpoint": "390c01494fb9260c6a06c26f7e5fa667cd4ee93e46b562b1fe34bba7c44286d3", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s9445a3f57fe8bd", "s52181b750b5b3b", "s51972e8ac79cc7", "sbc8e36698eb403", "s5d922eccc266ec", "scf0385959efbd2" ] }, { "age": 70, "six": 47.969166666666666, "five": 51.018, "tasks": { "blimp_filtered": 69.16, "supplement_filtered": 60.88, "ewok_filtered": 52.45, "entity_tracking": 20.26, "comps": 52.34, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 47.0 }, "checkpoint": "108c554eec9a3260f85204c6a13216e24eb00c1d2d505bd1d694ada022155f74", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sdafe7ce372c58a", "s2bd5f0e121d928", "s51972e8ac79cc7", "sbc8e36698eb403", "s5d922eccc266ec", "scf0385959efbd2" ] }, { "age": 80, "six": 48.86333333333334, "five": 51.006, "tasks": { "blimp_filtered": 69.49, "supplement_filtered": 59.67, "ewok_filtered": 52.01, "entity_tracking": 21.52, "comps": 52.34, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 53.0 }, "checkpoint": "3ea779b75b955b24d792d6923f1cf49577fd9db101eea49db1e96e2ddab4e66f", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4b6fddc55c1420", "s85bb71f38a509d", "s51972e8ac79cc7", "sbc8e36698eb403", "s5d922eccc266ec", "scf0385959efbd2" ] }, { "age": 90, "six": 48.024166666666666, "five": 50.602, "tasks": { "blimp_filtered": 68.53, "supplement_filtered": 58.46, "ewok_filtered": 51.92, "entity_tracking": 21.85, "comps": 52.25, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 46.0 }, "checkpoint": "cdb5206f9a7e813d615a59f071804255f7cfc61ad912c451d7a1089bae717846", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s0791b318131794", "sd9882c12ff7645", "s51972e8ac79cc7", "sbc8e36698eb403", "s5d922eccc266ec", "scf0385959efbd2" ] }, { "age": 100, "six": 47.93333333333333, "five": 50.29, "tasks": { "blimp_filtered": 68.84, "supplement_filtered": 59.88, "ewok_filtered": 51.36, "entity_tracking": 19.32, "comps": 52.05, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 49.0 }, "checkpoint": "55ccdf6c9960f7232e85af4c3aaa6eca3deb7727df10e0264e17228b7b420ff5", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sf42914cef0ef8a", "s2438dd7f12451d", "s51972e8ac79cc7", "sbc8e36698eb403", "s5d922eccc266ec", "scf0385959efbd2" ] } ], "peak": { "five": { "score": 51.128, "ages": [ 50 ], "rank": 31 }, "six": { "score": 48.86333333333334, "ages": [ 80 ], "rank": 47 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 38, "id": "run/IPRLM384_READER800_GATE420_DELTA_ROUTER125_SEED20260907_40M_20261022", "original_name": "IPRLM384_READER800_GATE420_DELTA_ROUTER125_SEED20260907_40M_20261022", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "Reader800/ゲート420＋差分Router125％", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All24same-seed40Mconditions;23existingcontrols;fullknowninventory;parentReader800gate420" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 141, "sources": [ "s70b55dbbbd3e1a", "sd5017e225e8bd1" ], "points": [ { "age": 40.0, "six": 47.9725, "five": 50.622, "tasks": { "blimp_filtered": 67.89, "supplement_filtered": 61.22, "ewok_filtered": 52.71, "entity_tracking": 18.81, "comps": 52.48, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 51.0 }, "checkpoint": "caadb2d8932529a3bed8db63fbe324da3fec8b108aeb9b58b938b84a9b16b997", "precision": "fp32", "verification": "primary", "sources": [ "scec6e0fa01cf59", "se86a654546165d" ] }, { "age": 70.0, "six": 48.85333333333333, "five": 51.394, "tasks": { "blimp_filtered": 71.19, "supplement_filtered": 61.3, "ewok_filtered": 51.83, "entity_tracking": 19.51, "comps": 53.14, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 49.0 }, "checkpoint": "5f68ae2860318de16c29397b8ff1183e556c52cd89ba64a9d7471d0e5b7dae6a", "precision": "fp32", "verification": "primary", "sources": [ "s53639bf3339c2b", "se86a654546165d" ] }, { "age": 80.0, "six": 48.1175, "five": 51.096, "tasks": { "blimp_filtered": 69.79, "supplement_filtered": 61.63, "ewok_filtered": 51.38, "entity_tracking": 19.73, "comps": 52.95, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 48.0 }, "checkpoint": "fef0fcd11d5e1e6317e256ba6a9b83ee984d0484632c969972d7dbf0dab3d06a", "precision": "fp32", "verification": "primary", "sources": [ "sbfc854929d2730", "se86a654546165d" ] } ], "peak": { "five": { "score": 51.394, "ages": [ 70.0 ], "rank": 13 }, "six": { "score": 48.85333333333333, "ages": [ 70.0 ], "rank": 48 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 39, "id": "run/IPRLM384_MOS_BRANCH_LR105_SEED20260907_80M_20261022", "original_name": "IPRLM384_MOS_BRANCH_LR105_SEED20260907_80M_20261022", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "MoS読み出し6行列LR＋5％（共同LR版ベース）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40/70/80M after80M training" ], "intermediate": "saved40/70M states evaluated after80M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000, 70000000, 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000, 70000000, 80000000 ], "timing": "after80M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All27LRconditions plus parent320 and full inventory; same-age jointAttentionReader parent" }, "milestones_words": [ 40000000, 70000000, 80000000 ], "checkpoints_words": [ 40000000, 70000000, 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 134, "sources": [ "sbbec738b3c3330", "s792d67daded52e" ], "points": [ { "age": 40.0, "six": 47.02333333333333, "five": 49.692, "tasks": { "blimp_filtered": 67.87, "supplement_filtered": 58.21, "ewok_filtered": 51.72, "entity_tracking": 18.5, "comps": 52.16, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 46.0 }, "checkpoint": "8890120395fe0279cd0a1dfd85ac350a62be5f007d7bc0afaca6b4653f926c3e", "precision": "fp32", "verification": "primary", "sources": [ "scc2d699fa18144", "s6d43fe0dd92cfb" ] }, { "age": 70.0, "six": 47.51, "five": 50.282, "tasks": { "blimp_filtered": 69.07, "supplement_filtered": 57.95, "ewok_filtered": 52.14, "entity_tracking": 20.15, "comps": 52.1, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 44.0 }, "checkpoint": "f35d3c3e24fa4389017139da20e154bd4785091e05da60be0cb2146846769bcf", "precision": "fp32", "verification": "primary", "sources": [ "se1081c3509296e", "s6d43fe0dd92cfb" ] }, { "age": 80.0, "six": 48.844166666666666, "five": 50.892, "tasks": { "blimp_filtered": 69.34, "supplement_filtered": 60.79, "ewok_filtered": 51.66, "entity_tracking": 20.4, "comps": 52.27, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 51.0 }, "checkpoint": "bf9e0f780a8d271f1f3e5b069d103422e835a283b246e801d51d4fdd9e094e25", "precision": "fp32", "verification": "primary", "sources": [ "sd61542e37ee5bd", "s6d43fe0dd92cfb" ] } ], "peak": { "five": { "score": 50.892, "ages": [ 80.0 ], "rank": 60 }, "six": { "score": 48.844166666666666, "ages": [ 80.0 ], "rank": 49 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 40, "id": "run/IPRLM320_FFN_ADDITIVE_BYPASS_SEED137_20261015", "original_name": "IPRLM320_FFN_ADDITIVE_BYPASS_SEED137_20261015", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at80M", "SyntaxGym31 and full PIQA1838", "frozen grounded2947" ], "intermediate": "No intermediate evaluation; train continuously to80M.", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 80000000 ], "timing": "after80M training", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Compare thirdseed80M to two existing additive seeds and historical parents; no matchedseed parent for the newseed." }, "milestones_words": [ 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 78, "sources": [ "s0c58006e3a1338", "s778e13c73a8199", "s6673c8e8d00507" ], "points": [ { "age": 40.0, "six": 47.30833333333333, "five": 49.928, "tasks": { "blimp_filtered": 69.54, "supplement_filtered": 57.87, "ewok_filtered": 51.44, "entity_tracking": 18.62, "comps": 52.17, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 49.0 }, "checkpoint": "2abb37fa2fdf989a8bd40e50e7aad682266b490687f9648d38544396c272ace5", "precision": "fp32", "verification": "primary", "sources": [ "s617da37f93627b", "s88c48189789e26" ] }, { "age": 70.0, "six": 48.76, "five": 51.188, "tasks": { "blimp_filtered": 70.35, "supplement_filtered": 59.97, "ewok_filtered": 51.9, "entity_tracking": 21.27, "comps": 52.45, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 48.0 }, "checkpoint": "a4e2b3c3e4a0c5692187d3a527d70cdf2526e16ec3fa26c86f7677d7790e92e1", "precision": "fp32", "verification": "primary", "sources": [ "safd9166364dc50", "s56a4d0229d85d3" ] }, { "age": 80.0, "six": 48.840833333333336, "five": 51.364, "tasks": { "blimp_filtered": 70.2, "supplement_filtered": 60.9, "ewok_filtered": 52.16, "entity_tracking": 20.99, "comps": 52.57, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 54.0 }, "checkpoint": "77b1157b4281c2f0ccb6187ca6fa99d510bfd8b0583a5a72d9772fc7ef08dc16", "precision": "fp32", "verification": "primary", "sources": [ "sc22c2720775619", "s3ce534941a2537" ] } ], "peak": { "five": { "score": 51.364, "ages": [ 80.0 ], "rank": 15 }, "six": { "score": 48.840833333333336, "ages": [ 80.0 ], "rank": 50 } }, "clear_points": [ { "number": 40, "age": 80, "seed": 137, "checkpoint_sha256": "77b1157b4281c2f0ccb6187ca6fa99d510bfd8b0583a5a72d9772fc7ef08dc16", "old_six": 48.840833333333336, "old_five": 51.364, "old_tasks": { "blimp_filtered": 70.2, "supplement_filtered": 60.9, "ewok_filtered": 52.16, "entity_tracking": 20.99, "comps": 52.57, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 54.0 }, "correct": 12245, "total": 17274, "accuracy": 70.88688201921963, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6946, "total": 8392, "ties": 0 }, "comps": { "correct": 3887, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 3, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 18, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 975, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 75, "total": 114, "ties": 0 }, "boolq": { "correct": 38, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 213, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 65.03279952600133, "checkpoint_rank": 10, "rank": 9 } ], "clear_peak": { "number": 40, "age": 80, "seed": 137, "checkpoint_sha256": "77b1157b4281c2f0ccb6187ca6fa99d510bfd8b0583a5a72d9772fc7ef08dc16", "old_six": 48.840833333333336, "old_five": 51.364, "old_tasks": { "blimp_filtered": 70.2, "supplement_filtered": 60.9, "ewok_filtered": 52.16, "entity_tracking": 20.99, "comps": 52.57, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 54.0 }, "correct": 12245, "total": 17274, "accuracy": 70.88688201921963, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6946, "total": 8392, "ties": 0 }, "comps": { "correct": 3887, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 3, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 18, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 975, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 75, "total": 114, "ties": 0 }, "boolq": { "correct": 38, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 213, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 65.03279952600133, "checkpoint_rank": 10, "rank": 9 }, "present": true }, { "number": 41, "id": "run/IPRLM384_FFN1408_LR10_MAJOR90_SEED20260907_40M_20261019", "original_name": "IPRLM384_FFN1408_LR10_MAJOR90_SEED20260907_40M_20261019", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "−10% 主要103行列だけ1割減", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "aftereachcase40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Three LR allocation cases and frozen baseline40M; full inventory; no final ranking from40M alone" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 113, "sources": [ "sb106aa1108a9f7", "s1b3c2c0ccd4277" ], "points": [ { "age": 40.0, "six": 48.165, "five": 50.662, "tasks": { "blimp_filtered": 69.24, "supplement_filtered": 59.41, "ewok_filtered": 52.89, "entity_tracking": 19.32, "comps": 52.45, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 50.0 }, "checkpoint": "0d66aded3ec322cd1bc1b604c01a07069d127b8a9c1dda54872dd494c9758422", "precision": "fp32", "verification": "primary", "sources": [ "s9b10408d4060e5", "sa3d65f2085be3b" ] }, { "age": 70.0, "six": 48.744166666666665, "five": 51.054, "tasks": { "blimp_filtered": 69.91, "supplement_filtered": 60.15, "ewok_filtered": 52.48, "entity_tracking": 20.54, "comps": 52.19, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 54.0 }, "checkpoint": "55ead07fde6e51efb1a101d84795efea2a5c305ec92c5c3f2f9dca400039712e", "precision": "fp32", "verification": "primary", "sources": [ "s1a755be0620529", "sa3d65f2085be3b" ] }, { "age": 80.0, "six": 48.8125, "five": 51.042, "tasks": { "blimp_filtered": 69.39, "supplement_filtered": 61.41, "ewok_filtered": 52.25, "entity_tracking": 19.77, "comps": 52.39, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 53.0 }, "checkpoint": "c78c4801e02e99e85f324c10c1cde4b13ef054503b923fb48a27d80073385b52", "precision": "fp32", "verification": "primary", "sources": [ "s225e11969ed1f6", "sa3d65f2085be3b" ] } ], "peak": { "five": { "score": 51.054, "ages": [ 70.0 ], "rank": 40 }, "six": { "score": 48.8125, "ages": [ 80.0 ], "rank": 52 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 42, "id": "run/Q3_MOS3_BRANCH123_WIDTH320_FFN1344_20261011", "original_name": "Q3_MOS3_BRANCH123_WIDTH320_FFN1344_20261011", "title": "320幅・FFN1344・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1344, "depth": 9, "parts": { "width": 320, "ffn": 1344, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 4.2 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1344, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes70/80/90/100M", "SyntaxGym31+PIQA1838 at70/80/90/100M", "frozen bAbI1947+ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Matched ages and observed peaks separately, including ordinary320 shared-last2 and all selected historical models" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 145, "sources": [ "sbc5cd5cf6e9e64", "s06ac8c61333822", "sa7c75f514bc0d8", "s977ebce27c96c1", "se8e87e43b59f5e", "sd6b8c8a767916d", "s97db4706830f42" ], "points": [ { "age": 70.0, "six": 48.81166666666667, "five": 50.85, "tasks": { "blimp_filtered": 70.9, "supplement_filtered": 59.94, "ewok_filtered": 51.31, "entity_tracking": 20.17, "comps": 51.93, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 52.0 }, "checkpoint": "4041d3dafd524671b6131479f77553f36712ec2e69eec471b97c8587176f62ec", "precision": "fp32", "verification": "primary", "sources": [ "se4e05addbeac5a", "sa4f0033921d6e5", "s2dbb86fe0a5f10", "sf8ed2173a9fb12", "sf01953fd8b6693", "sb9490b5c3d25fd", "s26cc7453f8ee04", "s9922595cb2ed8b" ] }, { "age": 80.0, "six": 48.784166666666664, "five": 50.902, "tasks": { "blimp_filtered": 69.36, "supplement_filtered": 61.3, "ewok_filtered": 51.0, "entity_tracking": 20.68, "comps": 52.17, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 56.0 }, "checkpoint": "3b5c7ec4a2fb502e47bb483c9c8e3defdaa34ca69fdd0b56d4eff4eeae60fa21", "precision": "fp32", "verification": "primary", "sources": [ "s567591b1299cc7", "saddaae5fd9abb3", "s05275ad5ae6579", "sf8ed2173a9fb12", "sf01953fd8b6693", "sb9490b5c3d25fd", "s26cc7453f8ee04", "s9922595cb2ed8b" ] }, { "age": 90.0, "six": 47.64833333333333, "five": 50.054, "tasks": { "blimp_filtered": 69.05, "supplement_filtered": 58.37, "ewok_filtered": 51.55, "entity_tracking": 19.15, "comps": 52.15, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 46.0 }, "checkpoint": "2e661bff2be54acb256dde0756d0dd6c4463b9552e30289d3cbc1e0799502784", "precision": "fp32", "verification": "primary", "sources": [ "s859b57fc747ecc", "s912ea9af3d4c6c", "se5e4af477e0f76", "sf8ed2173a9fb12", "sf01953fd8b6693", "sb9490b5c3d25fd", "s26cc7453f8ee04", "s9922595cb2ed8b" ] }, { "age": 100.0, "six": 48.18416666666667, "five": 50.594, "tasks": { "blimp_filtered": 69.23, "supplement_filtered": 60.44, "ewok_filtered": 51.86, "entity_tracking": 19.55, "comps": 51.89, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 48.0 }, "checkpoint": "ff528248a063fc772114511795b83148b849df9d329506a3d03f806402e8d94a", "precision": "fp32", "verification": "primary", "sources": [ "s5ea9aba2c6dd24", "se2582b6dc0f559", "s946c94fb07c9e6", "sf8ed2173a9fb12", "sf01953fd8b6693", "sb9490b5c3d25fd", "s26cc7453f8ee04", "s9922595cb2ed8b" ] } ], "peak": { "five": { "score": 50.902, "ages": [ 80.0 ], "rank": 57 }, "six": { "score": 48.81166666666667, "ages": [ 70.0 ], "rank": 53 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 43, "id": "run/IPRLM320_GATE_SELECTED_100M_20261011", "original_name": "IPRLM320_GATE_SELECTED_100M_20261011", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "prediction_residual_v1", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_init_only", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123_prediction_gate", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "initialization_layers": 9, "output_mixture_gate": "prediction_residual_v1", "mixture_gate_hidden": 320, "mixture_gate_statistics": "normalized_entropy_js_v1", "prediction_gate_correction_limit": 2.0, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes70/80/90/100M", "SyntaxGym31+PIQA1838 at70/80/90/100M", "bAbI1947+ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 TF32off exact prefix graph; actual9-layer320 prediction gate speed, score parity and all capacity buckets verified", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Matched ages and observed peaks separately, including ordinary320 shared-last2 and all selected historical models" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ], "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true, "compile_prediction_statistics": true } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 84, "sources": [ "sf40468e849b8e6", "s26c3641be21350", "sa43b832a0d9f1f", "s292050f602ba6c", "s4d3ef9b2f95901", "s31a99647ec3248" ], "points": [ { "age": 70.0, "six": 48.2975, "five": 50.624, "tasks": { "blimp_filtered": 70.72, "supplement_filtered": 58.48, "ewok_filtered": 51.16, "entity_tracking": 20.38, "comps": 52.38, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 51.0 }, "checkpoint": "4e89806c6278a02647b8846061016ba1986fc5a8f68d2751a8f693964999f031", "precision": "fp32", "verification": "primary", "sources": [ "sc7e9021f45eae5", "saa45b351912f1f", "s0c54df2d1f4c16", "sa6846897007fb3", "s10778c7d571950", "sfeb27c468974b0" ] }, { "age": 80.0, "six": 48.799166666666665, "five": 50.832, "tasks": { "blimp_filtered": 69.17, "supplement_filtered": 61.24, "ewok_filtered": 50.42, "entity_tracking": 21.17, "comps": 52.16, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 53.0 }, "checkpoint": "99b29662a9d4bfe4f595ac5e19947fd28f61507873e8dbd952757c02ef3369b5", "precision": "fp32", "verification": "primary", "sources": [ "s2af85e53b86f0a", "s96dafcd7ce4dcf", "s0c54df2d1f4c16", "sa6846897007fb3", "s10778c7d571950", "sfeb27c468974b0" ] }, { "age": 90.0, "six": 47.89, "five": 50.75, "tasks": { "blimp_filtered": 70.95, "supplement_filtered": 59.4, "ewok_filtered": 52.4, "entity_tracking": 19.09, "comps": 51.91, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 40.0 }, "checkpoint": "517b2cc080a27d1e46da70b262f91d554729435207a1cfafbe49f7ed2fefb192", "precision": "fp32", "verification": "primary", "sources": [ "s7ebb10a35730c2", "sefc5fb976ef31f", "s0c54df2d1f4c16", "sa6846897007fb3", "s10778c7d571950", "sfeb27c468974b0" ] }, { "age": 100.0, "six": 48.155, "five": 50.756, "tasks": { "blimp_filtered": 69.19, "supplement_filtered": 58.69, "ewok_filtered": 52.36, "entity_tracking": 20.63, "comps": 52.91, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 47.0 }, "checkpoint": "19dd4c36feba2142bc820575f16a49b41c0c89d79bc60a6c2065d6284d458fa3", "precision": "fp32", "verification": "primary", "sources": [ "se14388346467bc", "sb9bed6e20adcfc", "s0c54df2d1f4c16", "sa6846897007fb3", "s10778c7d571950", "sfeb27c468974b0" ] } ], "peak": { "five": { "score": 50.832, "ages": [ 80.0 ], "rank": 70 }, "six": { "score": 48.799166666666665, "ages": [ 80.0 ], "rank": 54 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 44, "id": "run/IPRLM320_ALL_BRANCH_SEGMENT_SELECTOR_FROM0_20261012", "original_name": "IPRLM320_ALL_BRANCH_SEGMENT_SELECTOR_FROM0_20261012", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": { "input": 2253, "hidden": 320, "outputs": 3, "eps": 1e-06, "normalization": "seven RMS-normalized320 blocks + seven raw RMS + three entropy + three JS", "initialization": "local RNG fan_in up; zero down; initial tie selects branch1 but all branches always receive direct answer loss", "feedback": "learned original MoS3 mixture; final output one selected component", "prediction": "argmax of standalone selector scores; one component distribution for eight predictions; reset per document", "segment_tokens": 8, "teacher": "same actual continuation, detached per-branch mean NLL", "loss": "equal mean of all three component next-token NLLs to body; detached segment preference loss to selector" }, "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "joint_mos3_all_branch_segment_selector", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "readout_gate": { "input": 2253, "hidden": 320, "outputs": 3, "eps": 1e-06, "normalization": "seven RMS-normalized320 blocks + seven raw RMS + three entropy + three JS", "initialization": "local RNG fan_in up; zero down; initial tie selects branch1 but all branches always receive direct answer loss", "feedback": "learned original MoS3 mixture; final output one selected component", "prediction": "argmax of standalone selector scores; one component distribution for eight predictions; reset per document", "segment_tokens": 8, "teacher": "same actual continuation, detached per-branch mean NLL", "loss": "equal mean of all three component next-token NLLs to body; detached segment preference loss to selector" }, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 exact recurrence with learned internal feedback and hard external selection", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true, "compile_readout_statistics": true }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M", "parent80M_six_axis": 49.8675, "parent80M_five_axis": 52.002, "no_endpoint_penalty": true }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": "resolved measured runtime selection", "compile_dynamic": "selector distribution statistics", "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 72, "sources": [ "s874b93623e7df9", "s1558da6b0ffe4d", "s15fc3e97461f6c", "s2f3b97b366d5c9", "s216b80f56b52c7", "s764e0b9d5075f5" ], "points": [ { "age": 70.0, "six": 48.471666666666664, "five": 50.942, "tasks": { "blimp_filtered": 70.11, "supplement_filtered": 59.56, "ewok_filtered": 52.25, "entity_tracking": 20.87, "comps": 51.92, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 47.0 }, "checkpoint": "5f1c8b794ed379541526fede9ecda782424e89332539a6ce4c9eedbcd250046f", "precision": "fp32", "verification": "primary", "sources": [ "s47be58182c9808", "s236e24be1f0bc0", "s95c4b34c68feed", "sb81fa8471dc3be", "s1f45f36c96ad08", "s512e2737ead9cd" ] }, { "age": 80.0, "six": 48.795, "five": 50.718, "tasks": { "blimp_filtered": 69.11, "supplement_filtered": 61.12, "ewok_filtered": 52.04, "entity_tracking": 19.1, "comps": 52.22, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 57.0 }, "checkpoint": "ed51e77f84b705bbf21e7767773d9710acb94e7aff4a11f3809eebf2c5cecb01", "precision": "fp32", "verification": "primary", "sources": [ "seea80584c389c4", "s6e59f246013ebd", "s95c4b34c68feed", "sb81fa8471dc3be", "s1f45f36c96ad08", "s512e2737ead9cd" ] }, { "age": 90.0, "six": 48.65, "five": 50.362, "tasks": { "blimp_filtered": 68.98, "supplement_filtered": 59.21, "ewok_filtered": 51.26, "entity_tracking": 20.55, "comps": 51.81, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 53.0 }, "checkpoint": "b444392062eedb9e9c5e8d34d1ad799481284736713291e15249b0138fb8ab51", "precision": "fp32", "verification": "primary", "sources": [ "s1c0213a5fc5818", "s3b7fd6aef3930b", "s95c4b34c68feed", "sb81fa8471dc3be", "s1f45f36c96ad08", "s512e2737ead9cd" ] }, { "age": 100.0, "six": 47.965, "five": 50.728, "tasks": { "blimp_filtered": 68.41, "supplement_filtered": 60.86, "ewok_filtered": 52.22, "entity_tracking": 19.78, "comps": 52.37, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 45.0 }, "checkpoint": "64a64fba96a19adb404f39a4a109f05e657dfd215277bea451f9e5c38d22fdd2", "precision": "fp32", "verification": "primary", "sources": [ "s6c7305c751172a", "s399c183a137b60", "s95c4b34c68feed", "sb81fa8471dc3be", "s1f45f36c96ad08", "s512e2737ead9cd" ] } ], "peak": { "five": { "score": 50.942, "ages": [ 70.0 ], "rank": 51 }, "six": { "score": 48.795, "ages": [ 80.0 ], "rank": 55 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 45, "id": "baby/vistarc_q3_mos2_received_value_gate_20261005", "original_name": "256・MoS2・3層参照＋受信ゲート", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos2_last3", "attention_received_gate": true, "ffn_input_output_gates": false, "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true }, "benchmark_words": [ 80000000, 100000000 ], "objective": { "metric": "maximum observed local six-axis", "train_words": 100000000, "save_words": [ 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "benchmark_words": [ 80000000, 100000000 ], "timing": "after100M training completes; serial", "parent_peak": 48.23583333333334, "historical_record": 49.03583333333333, "record_gap_from_parent": 0.79999999999999, "no_endpoint_penalty": true, "diagnostics": [ "same-checkpoint five-axis and seven component scores", "training loss", "candidate and reader usage", "mixture usage", "runtime" ], "comparison_scope": "Matched80/100M and parent complete60..100M; system comparison; runtime differences declared" } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260, "fused_mixture": "eager-equivalent Triton FP32 two-term logsumexp;native broadcast reduction backward" }, "condition_conflicts": [], "profile": 41, "sources": [ "s3e42225a43ede7", "s49550cf37cb854", "sd3c3dbcdc3001c", "s9a1fd1028741cd", "s60eced50e96689", "sff9cbc68753c43", "s5d196f25507b08", "s5862355568c021", "sf38cfbcb78b7e7", "s6af658d0990b1b" ], "points": [ { "age": 40.0, "six": 47.5425, "five": 49.924, "tasks": { "blimp_filtered": 68.24, "supplement_filtered": 58.34, "ewok_filtered": 51.34, "entity_tracking": 19.87, "comps": 51.83, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 47.0 }, "checkpoint": "2b22da46f602d4703cb3ef352e81c8ed96218c71de54e21b695f21f284731bfa", "precision": "fp32", "verification": "primary", "sources": [ "s838038ee7f5b58", "s6be6ab9f806269" ] }, { "age": 70, "six": 47.8775, "five": 50.626, "tasks": { "blimp_filtered": 69.91, "supplement_filtered": 59.31, "ewok_filtered": 50.62, "entity_tracking": 21.13, "comps": 52.16, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 44.0 }, "checkpoint": "2e8e175b9f0fc6c3491b6861a12001c01cddf28cbfe585c6a9699638248078fb", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s259a5234880910", "s755aa9c3f5c8de" ] }, { "age": 80, "six": 48.75666666666667, "five": 51.184, "tasks": { "blimp_filtered": 68.02, "supplement_filtered": 61.78, "ewok_filtered": 52.43, "entity_tracking": 21.12, "comps": 52.57, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 48.0 }, "checkpoint": "030468860fa2898d4eb30f7062771b40daa644f1f5a18320e256f9e4baffa169", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sce837c7ce53ae4", "s6680ff1f2f7de4", "s98c88207d7e3cd", "sd83e51e0f85986", "sc80584ed421714", "sfbeae53fd6ea8f", "s964aea9bcaac51", "sf450bbc5d751de", "s828b874667c5bc", "s2aa96c5e0ba863" ] }, { "age": 90, "six": 47.8625, "five": 50.896, "tasks": { "blimp_filtered": 67.88, "supplement_filtered": 60.36, "ewok_filtered": 52.81, "entity_tracking": 21.14, "comps": 52.29, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 45.0 }, "checkpoint": "beef8fd68d8148f0a317fecee1fc901ac66d9fd20ad99103b51cee7a37a4483d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc3725c9034122a", "s70df78ec4354ff" ] }, { "age": 100, "six": 47.454166666666666, "five": 50.712, "tasks": { "blimp_filtered": 68.54, "supplement_filtered": 58.82, "ewok_filtered": 52.55, "entity_tracking": 21.43, "comps": 52.22, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 40.0 }, "checkpoint": "dfbc9273317d76fb947a8795bbf9d50b1519feccd9c14bee1d9aadb832460a93", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s10370979f18bf1", "s6680ff1f2f7de4", "s98c88207d7e3cd", "sf13f90859332db", "sc80584ed421714", "sfbeae53fd6ea8f", "s964aea9bcaac51", "sf450bbc5d751de", "s828b874667c5bc", "s2aa96c5e0ba863" ] } ], "peak": { "five": { "score": 51.184, "ages": [ 80 ], "rank": 26 }, "six": { "score": 48.75666666666667, "ages": [ 80 ], "rank": 56 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 46, "id": "baby/B_READER_DEPTH9_WIDTH272_FAN_SCALED_100M", "original_name": "旧272・幅補正＋可変・距離補正あり", "title": "272幅・FFN960・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 272, "ffn": 960, "depth": 9, "parts": { "width": 272, "ffn": 960, "depth": 9, "attention": "8×34", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "学習する距離補正", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5294117647058822 }, "main_lr": [], "state": "採点済み・保存仕様の差あり", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 272, "layers": 9, "hidden": 960, "attention_heads": 8, "head_dim": 34, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "history_norm": "separate", "seed": 20260907, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_distance": "learned_log", "distance_init_min": 0.0625, "distance_init_max": 1.0, "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score60/70/80/90/100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [ "source:sd992bb83e06017", "source:s3e1667a0c1704c" ], "profile": 23, "sources": [ "s390d4d3d11aeb3", "sf4e5d1ab4c2762", "s8d27f6b1a97bc3", "sb350daee745878", "sd992bb83e06017", "sf91e10063ae5ee", "s74513a90a3397b", "s3e1667a0c1704c", "s100870ca6108ae", "s2e55832bc9dfc5" ], "points": [ { "age": 60, "six": 46.94833333333333, "five": 49.69, "tasks": { "blimp_filtered": 67.78, "supplement_filtered": 57.98, "ewok_filtered": 50.62, "entity_tracking": 19.81, "comps": 52.26, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 49.0 }, "checkpoint": "c7672b54b7ca96fe6cbe45c3378a04c2292166c973767a7a9169eef2991ed0b3", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s7964fdf3b36cc9", "s931b19f849313c", "s8d5efc0be121e4", "s9902064e8661b7", "sa1151ebeff3236", "s00eb5e6c328f70", "s6b90c1b26b415f", "s87506477c10a4f", "s8a42c649d4dbc4" ] }, { "age": 70, "six": 48.380833333333335, "five": 50.124, "tasks": { "blimp_filtered": 69.56, "supplement_filtered": 58.58, "ewok_filtered": 51.26, "entity_tracking": 19.39, "comps": 51.83, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 57.0 }, "checkpoint": "b9fdb858fd50ae14c56abf8e4eaed690f2b7f74c7ba369e6659b29c608c27d55", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s7964fdf3b36cc9", "s931b19f849313c", "s7d1cf014a02a18", "sa9bd0bdf1431cb", "sa1151ebeff3236", "s00eb5e6c328f70", "s6b90c1b26b415f", "s87506477c10a4f", "s8a42c649d4dbc4" ] }, { "age": 80, "six": 48.736666666666665, "five": 50.848, "tasks": { "blimp_filtered": 68.63, "supplement_filtered": 62.09, "ewok_filtered": 51.77, "entity_tracking": 19.77, "comps": 51.98, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 55.0 }, "checkpoint": "77a67017ffa6c5ea529f5bcbce4039eb3e29cabd81354c74e94fd0b95abc57b9", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s7964fdf3b36cc9", "s931b19f849313c", "s36ab5e3388c43f", "s8fd58e06e8d105", "sa1151ebeff3236", "s00eb5e6c328f70", "s6b90c1b26b415f", "s87506477c10a4f", "s8a42c649d4dbc4" ] }, { "age": 90, "six": 47.431666666666665, "five": 49.788, "tasks": { "blimp_filtered": 67.65, "supplement_filtered": 58.29, "ewok_filtered": 51.35, "entity_tracking": 19.6, "comps": 52.05, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 48.0 }, "checkpoint": "f98c6406acda702a59c0ae155a65d71a1c9a2596749d5f1d5efc9bcee061a8e7", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s7964fdf3b36cc9", "s931b19f849313c", "se50828a864155d", "sbe081ce8d9905b", "sa1151ebeff3236", "s00eb5e6c328f70", "s6b90c1b26b415f", "s87506477c10a4f", "s8a42c649d4dbc4" ] }, { "age": 100, "six": 48.026666666666664, "five": 50.102, "tasks": { "blimp_filtered": 69.18, "supplement_filtered": 58.31, "ewok_filtered": 51.53, "entity_tracking": 19.16, "comps": 52.33, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 52.0 }, "checkpoint": "0dad9c100816153473825c9ae8cab076bd202de0d9eb4aa2052cca1deba74e61", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s7964fdf3b36cc9", "s931b19f849313c", "s838195fc5e7fd4", "sebfd2a9d0831c2", "sa1151ebeff3236", "s00eb5e6c328f70", "s6b90c1b26b415f", "s87506477c10a4f", "s8a42c649d4dbc4" ] } ], "peak": { "five": { "score": 50.848, "ages": [ 80 ], "rank": 68 }, "six": { "score": 48.736666666666665, "ages": [ 80 ], "rank": 57 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 47, "id": "run/IPRLM320_JOINT_MIXTURE_READOUT_FROM0_20261012", "original_name": "IPRLM320_JOINT_MIXTURE_READOUT_FROM0_20261012", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": { "input": 2253, "hidden": 320, "outputs": 3, "eps": 1e-06, "normalization": "seven RMS-normalized320 blocks + seven raw RMS + three entropy + three JS", "initialization": "local seeded fanin up; zero down and biases; reproduces parent0M output initially", "feedback": "original MoS3 mixture; original body/mixing parameters learn jointly from0", "prediction": "softmax(original mixing scores + residual readout scores); per-token probability mixture", "gradient": "gate input statistics and features detached; mixed NLL updates both original model and readout gate" }, "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "joint_parent_mos3_readout_gate", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "readout_gate": { "input": 2253, "hidden": 320, "outputs": 3, "eps": 1e-06, "normalization": "seven RMS-normalized320 blocks + seven raw RMS + three entropy + three JS", "initialization": "local seeded fanin up; zero down and biases; reproduces parent0M output initially", "feedback": "original MoS3 mixture; original body/mixing parameters learn jointly from0", "prediction": "softmax(original mixing scores + residual readout scores); per-token probability mixture", "gradient": "gate input statistics and features detached; mixed NLL updates both original model and readout gate" }, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 exact recurrence; original mixed internalfeedback; extra residual mixing finaloutput", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true, "compile_readout_statistics": true }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M", "parent80M_six_axis": 49.8675, "parent80M_five_axis": 52.002, "no_endpoint_penalty": true }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ], "cancelled_comparators": [ { "run": "IPRLM320_SEGMENT_SELECTOR_FROM0_20261012", "reason": "User stopped at40.763785M before scheduled benchmarks; no scores available and no additional scoring requested" } ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": "resolved measured runtime selection", "compile_dynamic": "readout distribution statistics", "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 86, "sources": [ "s4cce8c4ea2c32f", "s556a6ecc207377", "s87907721c1ad5e", "sdfe0a4d2adfe0d", "se5a8c2b4f87197", "s930cbfd4fc8f92" ], "points": [ { "age": 40.0, "six": 46.81666666666667, "five": 49.844, "tasks": { "blimp_filtered": 69.95, "supplement_filtered": 56.99, "ewok_filtered": 50.69, "entity_tracking": 19.37, "comps": 52.22, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 42.0 }, "checkpoint": "badc1fd36e93eece01d2384391522fbbe38cee6cdee46a4d6a89dea3a295e899", "precision": "fp32", "verification": "primary", "sources": [ "s60cfd90263731d", "sffff1eaad5c508" ] }, { "age": 70.0, "six": 48.406666666666666, "five": 50.47, "tasks": { "blimp_filtered": 70.03, "supplement_filtered": 57.42, "ewok_filtered": 51.19, "entity_tracking": 21.29, "comps": 52.42, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 49.0 }, "checkpoint": "47075404edc3d4ff2a8fbefcaa77a6f336ff6a5c05bb183a73a12f988b210ed1", "precision": "fp32", "verification": "primary", "sources": [ "sf5fe64815ca642", "s81e0c23aa51d40", "s7baf6f6c051e4d", "s080b3f6d763052", "sd3733dbc542ccd", "sadf9629ca9cacf" ] }, { "age": 80.0, "six": 48.71666666666667, "five": 50.63, "tasks": { "blimp_filtered": 69.42, "supplement_filtered": 58.91, "ewok_filtered": 51.73, "entity_tracking": 20.6, "comps": 52.49, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 55.0 }, "checkpoint": "3d0d044da5b6007edaf8bb0f2cadc46f824f9432a32c6232f702a52946dbdaa6", "precision": "fp32", "verification": "primary", "sources": [ "scaf765efea7c7e", "sf6c44534d6b2db", "s7baf6f6c051e4d", "s080b3f6d763052", "sd3733dbc542ccd", "sadf9629ca9cacf" ] }, { "age": 90.0, "six": 47.928333333333335, "five": 50.484, "tasks": { "blimp_filtered": 70.2, "supplement_filtered": 57.09, "ewok_filtered": 51.96, "entity_tracking": 20.9, "comps": 52.27, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 47.0 }, "checkpoint": "132e62938ead5655c53c9f1e4c289afd0d5af448a5b13791a9f6be33cc11d650", "precision": "fp32", "verification": "primary", "sources": [ "se77b050bfff3c7", "sc35aaa0cc0ca5d", "s7baf6f6c051e4d", "s080b3f6d763052", "sd3733dbc542ccd", "sadf9629ca9cacf" ] }, { "age": 100.0, "six": 48.096666666666664, "five": 50.286, "tasks": { "blimp_filtered": 69.06, "supplement_filtered": 56.84, "ewok_filtered": 51.21, "entity_tracking": 21.89, "comps": 52.43, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "c88d20c2245bdf4b411085bfb583e7e7ecaca2ae2c7101bff7df942cdd5520a9", "precision": "fp32", "verification": "primary", "sources": [ "s059f5af1afd0ba", "s9c7d99aea02c32", "s7baf6f6c051e4d", "s080b3f6d763052", "sd3733dbc542ccd", "sadf9629ca9cacf" ] } ], "peak": { "five": { "score": 50.63, "ages": [ 80.0 ], "rank": 92 }, "six": { "score": 48.71666666666667, "ages": [ 80.0 ], "rank": 59 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 48, "id": "baby/B_READER_DEPTH9_LEARNED_CONTINUOUS_100M", "original_name": "旧256・9層・可変ノブ＋距離補正", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "学習する距離補正", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_distance": "learned_log", "distance_init_min": 0.0625, "distance_init_max": 1.0, "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "Parent80M score is reused;100M score added after completion", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 18, "sources": [ "s243d89e4c2e29a", "s708a0f7fdc9455", "s5177f0a8d65eb2", "scfed5e931d8e13", "sd286dca251a9fd", "sbd9049fc3e496a", "see1b1d0ca0e886", "sbc56f9239d2821", "s3a8a6067342269", "sa1ff8bdd259c2c", "s2b39f1560eb5ba" ], "points": [ { "age": 80, "six": 48.715, "five": 50.934, "tasks": { "blimp_filtered": 68.64, "supplement_filtered": 60.91, "ewok_filtered": 52.51, "entity_tracking": 20.3, "comps": 52.31, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 50.0 }, "checkpoint": "4b6f80993aec7136a35db79cdb868b27b44e6d96a9c360e8c49d9634ab2f29a5", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s2d45e8d7d73325", "seb6033d89bf7bb", "sa92b4934ef3ef7", "sbcb5a233467a10", "s29981d15293c9d", "sdd1b5e9dca361a", "se1999ca80c5f5c", "s9feaeb7fcdeffb", "s3110f210789fd1", "s0787ec8ee10976", "sb0980a3feaba24", "s0d86771572bf06", "se58eb282e52842" ] }, { "age": 100, "six": 47.594166666666666, "five": 50.092, "tasks": { "blimp_filtered": 68.76, "supplement_filtered": 58.71, "ewok_filtered": 51.24, "entity_tracking": 19.43, "comps": 52.32, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 44.0 }, "checkpoint": "60e62be6aa09d94efedd1633d012706fb5b99b15cd0876338d02a47c19687279", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s2d45e8d7d73325", "seb6033d89bf7bb", "scc40cadfd2f3c0", "sde27882e4d08e4", "s9feaeb7fcdeffb", "s0787ec8ee10976", "s0d86771572bf06" ] } ], "peak": { "five": { "score": 50.934, "ages": [ 80 ], "rank": 53 }, "six": { "score": 48.715, "ages": [ 80 ], "rank": 60 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 49, "id": "baby/vistarc_q3_mos3_full_delta9876_width320_20261010", "original_name": "320・MoS3・全幅差分9−8/8−7/7−6", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別2/2/2層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9・8層 / 8・7層 / 7・6層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9, 8 ], [ 8, 7 ], [ 7, 6 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_full_delta9876", "attention_received_gate": true, "output_components": 3, "output_state_layers": 4, "reference_attention_width": 256, "output_branch_layers": [ [ 9, 8 ], [ 8, 7 ], [ 7, 6 ] ], "output_branch_operator": "signed_difference", "output_difference_width": 320, "output_raw_state_bypass": false, "output_mixing_input": "concatenated_three_full_width_difference_contexts", "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 TF32off exact recurrent signed-difference outputs; GPU parity, real-text throughput and total graph cache capacity checked.", "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Full delta320 against nested320, separate987320 and compressed64FiLM320 at matched ages and observed peaks, with historical references." }, "milestones_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 46, "sources": [ "s3e1c391af97d22", "s49e32fc0ece13a", "s5e8a7d88dc06cf", "sfbcb5675b40c18", "s59ba4b23820e2a", "sd6688ac75669bd", "se94bf35077557d" ], "points": [ { "age": 70, "six": 48.715, "five": 50.634, "tasks": { "blimp_filtered": 69.07, "supplement_filtered": 60.17, "ewok_filtered": 51.26, "entity_tracking": 20.61, "comps": 52.06, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 53.0 }, "checkpoint": "ec1dbc4062e48e085090223d771cc2f012e9af01656e0caa01ecfb9f14fc16b3", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s85b2b8f89a7823", "sf94595c35d4cd4", "s6e1fc629e012c4", "s2313a015803315", "s9363ca012e5a1d", "sed7685ca3333b1" ] }, { "age": 80, "six": 48.278333333333336, "five": 50.598, "tasks": { "blimp_filtered": 69.53, "supplement_filtered": 61.97, "ewok_filtered": 50.16, "entity_tracking": 19.1, "comps": 52.23, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 52.0 }, "checkpoint": "2657204fd4ff3181a19d91d45c5f4b05d571d26fdc4bd7d478c87c672fb5a454", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sb924cd5220998e", "s4b546e02e49c07", "s6e1fc629e012c4", "s2313a015803315", "s9363ca012e5a1d", "sed7685ca3333b1" ] }, { "age": 90, "six": 47.788333333333334, "five": 50.41, "tasks": { "blimp_filtered": 68.61, "supplement_filtered": 59.72, "ewok_filtered": 51.16, "entity_tracking": 20.09, "comps": 52.47, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 48.0 }, "checkpoint": "8857cb42d9d19eb962aad4b3c4be13b408ab8916f5e1f07465d6283babfeabc1", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s17301e8c4200a5", "sc85d43789d9174", "s6e1fc629e012c4", "s2313a015803315", "s9363ca012e5a1d", "sed7685ca3333b1" ] } ], "peak": { "five": { "score": 50.634, "ages": [ 70 ], "rank": 90 }, "six": { "score": 48.715, "ages": [ 70 ], "rank": 60 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 50, "id": "baby/vistarc_q3_mos4_received_value_gate_20261006", "original_name": "256・MoS4・3層混合", "title": "256幅・FFN896・9層", "topology": "MoS4", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS4", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 4, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos4_last3", "attention_received_gate": true, "output_components": 4, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS4 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 80000000, 100000000 ], "objective": { "metric": "maximum observed local six-axis", "train_words": 100000000, "save_words": [ 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "benchmark_words": [ 80000000, 100000000 ], "timing": "after100M training completes; serial", "historical_record": 49.03583333333333, "no_endpoint_penalty": true, "diagnostics": [ "same-checkpoint five-axis and seven component scores", "training loss", "candidate and reader usage", "mixture usage", "runtime" ], "comparison_scope": "Original MoS3 received-gate versus MoS4 received-gate; completed same80/100M points; no FFN gates" } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 60, "sources": [ "s4001b12b6f13d8", "s100d81cb8d541b", "s25efd99590c448", "sf245b0f0227084", "s753dc4fb5291a7", "s0a3e660c257260", "sefbd579958a1b8" ], "points": [ { "age": 40.0, "six": 47.49, "five": 49.352, "tasks": { "blimp_filtered": 68.01, "supplement_filtered": 56.3, "ewok_filtered": 51.18, "entity_tracking": 19.4, "comps": 51.87, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 55.0 }, "checkpoint": "89cafb889a4aeea38fb46aede60a93682460d9a7982ec127f2afb22b0e44e416", "precision": "fp32", "verification": "primary", "sources": [ "scb2096faf02780", "s530d829cf497aa" ] }, { "age": 70, "six": 48.276666666666664, "five": 50.69, "tasks": { "blimp_filtered": 70.32, "supplement_filtered": 58.64, "ewok_filtered": 52.33, "entity_tracking": 20.03, "comps": 52.13, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 53.0 }, "checkpoint": "11f1a4a10e45d889ca78bfd10b654c57f0947995bd849d95eccf37bca6feaf8e", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s30d622b2f74577", "s4f456f6d5f1392" ] }, { "age": 80, "six": 48.50083333333333, "five": 50.768, "tasks": { "blimp_filtered": 69.18, "supplement_filtered": 59.12, "ewok_filtered": 52.0, "entity_tracking": 20.95, "comps": 52.59, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "checkpoint": "cff3e9d67bd1ae0225f675a4bf8c7ba34747142a6cdaccffe6e067edfb66a62d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s90055a016b87a8", "s25ae3f80c39006", "s628a736b4264e6", "s05ac9c4ef0ea93", "s4936e932193a62", "sd168b41a430b10", "sb09f66ccda459d" ] }, { "age": 90, "six": 47.89833333333333, "five": 50.242, "tasks": { "blimp_filtered": 68.8, "supplement_filtered": 58.93, "ewok_filtered": 51.42, "entity_tracking": 19.51, "comps": 52.55, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 51.0 }, "checkpoint": "96659eb1084e7d82e2dd6d236ea2e71bd86a920acb2375269be080995cb8b314", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "se7c764158bc03a", "s0ca1599f0034f9" ] }, { "age": 100, "six": 48.69083333333333, "five": 50.896, "tasks": { "blimp_filtered": 69.03, "supplement_filtered": 60.77, "ewok_filtered": 51.03, "entity_tracking": 20.81, "comps": 52.84, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 53.0 }, "checkpoint": "f674f0ff2f9cee7cc62c4840a261ab6d942aacb456e06ce6100838b0b7b9c05a", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc4aea1e91f0a80", "s25ae3f80c39006", "s0360905b18597a", "s05ac9c4ef0ea93", "s4936e932193a62", "sd168b41a430b10", "sb09f66ccda459d" ] } ], "peak": { "five": { "score": 50.896, "ages": [ 100 ], "rank": 58 }, "six": { "score": 48.69083333333333, "ages": [ 100 ], "rank": 63 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 51, "id": "run/IPRLM384_BEST1344_LAST2_B_LR_SEED20260907_80M_20261024", "original_name": "IPRLM384_BEST1344_LAST2_B_LR_SEED20260907_80M_20261024", "title": "384幅・FFN1344・9層", "topology": "MoS3・共通2層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1344, "depth": 9, "parts": { "width": 384, "ffn": 1344, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1344, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last2", "attention_received_gate": true, "output_components": 3, "output_state_layers": 2, "reference_attention_width": 256, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six-axis at40/70/80M after80M training" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3_LAST2 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 65536, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000, 70000000, 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "five_axis_and_all_tasks", "benchmark_words": [ 40000000, 70000000, 80000000 ], "no_endpoint_penalty": true, "comparisons": "structural highest384 parent, B LR donor, all historical conditions and matched ages" }, "milestones_words": [ 40000000, 70000000, 80000000 ], "checkpoints_words": [ 40000000, 70000000, 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 103, "sources": [ "s7755cea57a66ad" ], "points": [ { "age": 40.0, "six": 47.56916666666667, "five": 50.038, "tasks": { "blimp_filtered": 68.54, "supplement_filtered": 57.79, "ewok_filtered": 51.14, "entity_tracking": 20.18, "comps": 52.54, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 52.0 }, "checkpoint": "16c8c81a7a9ef3122c1d504aaeaffda1c0177444c97de4ca7ba653fddf3230f7", "precision": "fp32", "verification": "primary", "sources": [ "s8ac6c255436035" ] }, { "age": 70.0, "six": 48.35333333333333, "five": 50.588, "tasks": { "blimp_filtered": 69.22, "supplement_filtered": 59.7, "ewok_filtered": 51.67, "entity_tracking": 19.99, "comps": 52.36, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 53.0 }, "checkpoint": "a3c5ca63fc70863b04f4484cb7e369a2b82fc22874f010e4a9f8490bf729c629", "precision": "fp32", "verification": "primary", "sources": [ "s3a8bbd82d53a3d" ] }, { "age": 80.0, "six": 48.67166666666667, "five": 50.67, "tasks": { "blimp_filtered": 68.9, "supplement_filtered": 59.63, "ewok_filtered": 52.55, "entity_tracking": 20.06, "comps": 52.21, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 56.0 }, "checkpoint": "2163a39facc5ddb04938b20aec6922d4caddb43d62ccfbe3ef04185e9171b7ef", "precision": "fp32", "verification": "primary", "sources": [ "s07c6a3a210debd" ] } ], "peak": { "five": { "score": 50.67, "ages": [ 80.0 ], "rank": 86 }, "six": { "score": 48.67166666666667, "ages": [ 80.0 ], "rank": 64 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 52, "id": "baby/vistarc_q3_mos3_received_ffn_gates_20261006", "original_name": "256・MoS3・3層混合＋FFN両ゲート", "title": "256幅・FFN896・9層", "topology": "MoS3", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "あり", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last3", "attention_received_gate": true, "output_components": 3, "ffn_input_output_gates": true, "ffn_gate_rank": 32 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 80000000, 100000000 ], "objective": { "metric": "maximum observed local six-axis", "train_words": 100000000, "save_words": [ 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "benchmark_words": [ 80000000, 100000000 ], "timing": "after100M training completes; serial", "historical_record": 49.03583333333333, "no_endpoint_penalty": true, "diagnostics": [ "same-checkpoint five-axis and seven component scores", "training loss", "candidate and reader usage", "mixture usage", "runtime" ], "comparison_scope": "Combined model versus MoS3 and single-output FFN-gate parents. No score inferred from sum of component improvements." } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 56, "sources": [ "s541b28271ffb70", "sfecae0c7fda4e3", "s51b51d1fb3bf8c", "sea1c06e59c052b", "s5160fb5c758c42", "s12ebd433380b1a", "sfb4deed4c1a96d" ], "points": [ { "age": 70, "six": 48.56333333333333, "five": 50.846, "tasks": { "blimp_filtered": 70.13, "supplement_filtered": 58.74, "ewok_filtered": 51.39, "entity_tracking": 21.23, "comps": 52.74, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "fa1e72c656828fc8d3c523d5ae79b866fd9ef8e62c8af435a4589400a726a812", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s744bf8fdf03a43", "sc32176236471e3" ] }, { "age": 80, "six": 48.66166666666667, "five": 51.152, "tasks": { "blimp_filtered": 68.95, "supplement_filtered": 61.86, "ewok_filtered": 50.85, "entity_tracking": 21.23, "comps": 52.87, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 53.0 }, "checkpoint": "ad1b77b7279391cc840e65c776ffdbae087cde7e5cbb19bae4f3ef09ce09d69e", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1668588df057f4", "s133f7650a3dbb6", "sf61a16d01f4ea6", "sd00d0009cb55fb", "sf0735fcd94efe8", "s93b873a6740ab3", "sab648000fb84b2" ] }, { "age": 90, "six": 48.13333333333333, "five": 50.43, "tasks": { "blimp_filtered": 68.62, "supplement_filtered": 58.13, "ewok_filtered": 51.46, "entity_tracking": 21.02, "comps": 52.92, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "7a6f52b750f0d5cc86e2b735be9e02c11561a9fab2f11da3d08780e42f824bdd", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sccf22fb6e6adb4", "s59bd55da2048e4" ] }, { "age": 100, "six": 48.083333333333336, "five": 50.864, "tasks": { "blimp_filtered": 69.08, "supplement_filtered": 60.08, "ewok_filtered": 52.33, "entity_tracking": 19.99, "comps": 52.84, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 47.0 }, "checkpoint": "8dcbbb3a5c9387ec772e92fb7c21281f9bc1b0fe462b83a370c4b52d52c9ea5b", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s16ca15a3c283bf", "s133f7650a3dbb6", "sc8a32352abf2fc", "sd00d0009cb55fb", "sf0735fcd94efe8", "s93b873a6740ab3", "sab648000fb84b2" ] } ], "peak": { "five": { "score": 51.152, "ages": [ 80 ], "rank": 28 }, "six": { "score": 48.66166666666667, "ages": [ 80 ], "rank": 65 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 53, "id": "run/IPRLM384_LATE_READOUT_SWITCH_70TO80M_SEED20260907_20261023", "original_name": "IPRLM384_LATE_READOUT_SWITCH_70TO80M_SEED20260907_20261023", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "384・70M以降のみ読出し342／混合360", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at80M" ], "intermediate": "40M/70M are inherited parent states; no new scoring", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 80000000 ], "timing": "after80M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All29priorLRconditions plus new late-switch; parent320 and fullinventory" }, "milestones_words": [ 80000000 ], "checkpoints_words": [ 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 111, "sources": [ "sc4ea967c4cad69", "s39cd2df050a188" ], "points": [ { "age": 80.0, "six": 48.64833333333333, "five": 50.942, "tasks": { "blimp_filtered": 69.9, "supplement_filtered": 59.21, "ewok_filtered": 51.3, "entity_tracking": 21.97, "comps": 52.33, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 53.0 }, "checkpoint": "f31a412f8ad898ad3b1d6accacb23b21cc322f55fb1c554201f5254b0593635a", "precision": "fp32", "verification": "primary", "sources": [ "s1f13f5032895fe", "s342ac49ae678c0" ] } ], "peak": { "five": { "score": 50.942, "ages": [ 80.0 ], "rank": 51 }, "six": { "score": 48.64833333333333, "ages": [ 80.0 ], "rank": 66 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 54, "id": "run/IPRLM336_OLD_CANDIDATE_READER48_RECEIVED_MOS3_NO_FFN_ADD_SEED20260907_20261018", "original_name": "IPRLM336_OLD_CANDIDATE_READER48_RECEIVED_MOS3_NO_FFN_ADD_SEED20260907_20261018", "title": "336幅・FFN1232・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at70M and80M" ], "intermediate": "After80M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000 ], "timing": "after80M", "no_endpoint_penalty": true, "comparison_scope": "Same-seed70/80M old-candidate additive parent336 and completed new-candidate Reader48 MoS3 without FFN addition." }, "milestones_words": [ 70000000, 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 94, "sources": [ "s3d396bba9d7652", "sbc2d7a83c0ddc1", "s085fb6bab6d62c" ], "points": [ { "age": 70.0, "six": 48.6175, "five": 50.614, "tasks": { "blimp_filtered": 70.31, "supplement_filtered": 59.0, "ewok_filtered": 51.05, "entity_tracking": 20.36, "comps": 52.35, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 53.0 }, "checkpoint": "fc7b8e660d9469d89792e82d5d8ef8026927086c488c86a6523562f567e505a4", "precision": "fp32", "verification": "primary", "sources": [ "sb90d9e29cedac8", "s94c9630c3b0552" ] }, { "age": 80.0, "six": 48.5525, "five": 50.93, "tasks": { "blimp_filtered": 69.16, "supplement_filtered": 61.56, "ewok_filtered": 49.8, "entity_tracking": 21.52, "comps": 52.61, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 51.0 }, "checkpoint": "d712df924db5bec0c59fbe79f997d2e1679ccea22820f05dd557f81febe7ea57", "precision": "fp32", "verification": "primary", "sources": [ "s4cd10a5b87a3a0", "s0dad6eac882e87" ] } ], "peak": { "five": { "score": 50.93, "ages": [ 80.0 ], "rank": 54 }, "six": { "score": 48.6175, "ages": [ 70.0 ], "rank": 67 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 55, "id": "run/IPRLM320_MOS3_INDEPENDENT_VOCAB_20261013", "original_name": "IPRLM320_MOS3_INDEPENDENT_VOCAB_20261013", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123_independent_vocab", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000 at40/70/80/90/100M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for independent vocabulary head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false, "vocabulary_backend": "loop" }, "benchmark_words": [ 40000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Independent vocabulary matrices versus parent320; matched40/70/80/90/100M local scores and observed peaks. Performance not yet measured." }, "milestones_words": [ 40000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "After100M training completes" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 87, "sources": [ "s094fa2304ae883", "sac7be605e33105", "s5802a1339ab791" ], "points": [ { "age": 40.0, "six": 47.579166666666666, "five": 49.974, "tasks": { "blimp_filtered": 68.39, "supplement_filtered": 57.8, "ewok_filtered": 51.37, "entity_tracking": 20.2, "comps": 52.11, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 45.0 }, "checkpoint": "21a7ec36bfb9fd9d4e553f38bbc6b43ed3b2e4a887104eaa629c248339911648", "precision": "fp32", "verification": "primary", "sources": [ "sc249daa80a4c5b", "se7843dc09d718a", "s55450e63f36510", "sb1cfb723658a52" ] }, { "age": 70.0, "six": 48.593333333333334, "five": 50.882, "tasks": { "blimp_filtered": 69.35, "supplement_filtered": 60.11, "ewok_filtered": 51.11, "entity_tracking": 21.47, "comps": 52.37, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "ebdf61861753698dbc2c22fa1c3785e460e61aef12155f3f4026f5347f556009", "precision": "fp32", "verification": "primary", "sources": [ "s536937306bc21d", "s13f706311b572e", "s55450e63f36510", "sb1cfb723658a52" ] }, { "age": 80.0, "six": 48.134166666666665, "five": 50.228, "tasks": { "blimp_filtered": 68.47, "supplement_filtered": 59.7, "ewok_filtered": 51.3, "entity_tracking": 19.81, "comps": 51.86, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 53.0 }, "checkpoint": "f21f69950be7398d4b6adbe59710269d38470865c6ae93216318c790f9f69d69", "precision": "fp32", "verification": "primary", "sources": [ "s529c683f7240d4", "s9ded422bdc5044", "s55450e63f36510", "sb1cfb723658a52" ] }, { "age": 90.0, "six": 47.964166666666664, "five": 50.03, "tasks": { "blimp_filtered": 68.42, "supplement_filtered": 58.4, "ewok_filtered": 50.05, "entity_tracking": 21.11, "comps": 52.17, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 51.0 }, "checkpoint": "75b79e3d23557aa59ed21a66845427b8fb3a9b684a0f0756daaa5d338354d4bd", "precision": "fp32", "verification": "primary", "sources": [ "s31c8f7a4569d63", "s8881d6bafdbafa", "s55450e63f36510", "sb1cfb723658a52" ] }, { "age": 100.0, "six": 48.32333333333333, "five": 50.658, "tasks": { "blimp_filtered": 68.29, "supplement_filtered": 59.82, "ewok_filtered": 52.92, "entity_tracking": 19.97, "comps": 52.29, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "b3f76f68c103f9ea0f81531f6b2a82ec3ac320b3f7afc69fdb2866101e6524d8", "precision": "fp32", "verification": "primary", "sources": [ "s63d1c7aa99684b", "sdbead185ec943e", "s55450e63f36510", "sb1cfb723658a52" ] } ], "peak": { "five": { "score": 50.882, "ages": [ 70.0 ], "rank": 63 }, "six": { "score": 48.593333333333334, "ages": [ 70.0 ], "rank": 70 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 56, "id": "baby/Q3_READER48_BALANCED", "original_name": "Q3・Reader48 Balanced", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 31, "sources": [ "s10fbd748ab877c", "sfdf9ed2203fb25", "se9f61896d8d65b", "sc37a75a18832ff", "s1c418385541984", "s0c83427fe25fb3" ], "points": [ { "age": 80, "six": 48.575833333333335, "five": 50.864, "tasks": { "blimp_filtered": 68.42, "supplement_filtered": 58.94, "ewok_filtered": 53.75, "entity_tracking": 20.64, "comps": 52.57, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 50.0 }, "checkpoint": "0273ac9b6600fd5415e0d3dd026b7f2ba7faeeffbdb1f31f397efe665d00e918", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s36d6c100b4e7be", "s99ce7598e7dd13", "s685d4316a782ef", "se70fa21f8b0676", "s7e808b5a16e609", "sd7fac3eac2ff07", "s96b3ef57680a72", "s8683f2226f6e45" ] }, { "age": 100, "six": 47.39833333333333, "five": 50.142, "tasks": { "blimp_filtered": 67.59, "supplement_filtered": 57.85, "ewok_filtered": 52.63, "entity_tracking": 19.89, "comps": 52.75, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 46.0 }, "checkpoint": "7feef037d4fe53c05ee07fa4f2fd7e9430ef6c60026a678fb77f2d3853ad6c54", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s36d6c100b4e7be", "s99ce7598e7dd13", "sa48dab99f8d855", "s13cafa458fb96f", "s7e808b5a16e609", "sd7fac3eac2ff07", "s96b3ef57680a72", "s8683f2226f6e45" ] } ], "peak": { "five": { "score": 50.864, "ages": [ 80 ], "rank": 66 }, "six": { "score": 48.575833333333335, "ages": [ 80 ], "rank": 71 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 57, "id": "run/IPRLM_MOS3_256_LAST3_SEED731946281_100M_20261014", "original_name": "IPRLM_MOS3_256_LAST3_SEED731946281_100M_20261014", "title": "256幅・FFN896・9層", "topology": "MoS3", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last3", "attention_received_gate": true, "output_components": 3, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes" ], "intermediate": "After100M train:70/80/90/100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "selection_metric": "five_axis", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "no_endpoint_penalty": true }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 154, "sources": [ "sc85f362b1be87d", "s422c0642b38f54", "sdeae0e10d3977d" ], "points": [ { "age": 70.0, "six": 48.32666666666667, "five": 50.468, "tasks": { "blimp_filtered": 69.76, "supplement_filtered": 59.48, "ewok_filtered": 51.06, "entity_tracking": 20.07, "comps": 51.97, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 50.0 }, "checkpoint": "84876d6c2bf38e4db53a98f09fad8f711b4609c2d2cc9a573cb7e20b83d0f65f", "precision": "fp32", "verification": "primary", "sources": [ "saa18c1d7ee6788", "s38121ce3d3494e" ] }, { "age": 80.0, "six": 48.545, "five": 50.912, "tasks": { "blimp_filtered": 69.23, "supplement_filtered": 61.32, "ewok_filtered": 51.85, "entity_tracking": 19.93, "comps": 52.23, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 54.0 }, "checkpoint": "a3dff7f53eec272a3e72bf9bdb799e1cdfe7d2b7769b0de423d50fecb1966795", "precision": "fp32", "verification": "primary", "sources": [ "scefd8fa8a8e44b", "se0032aa5fe5cca" ] }, { "age": 90.0, "six": 47.59916666666667, "five": 50.286, "tasks": { "blimp_filtered": 68.56, "supplement_filtered": 58.41, "ewok_filtered": 51.12, "entity_tracking": 21.21, "comps": 52.13, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "51f02d54121c345bdfdf1b5c855cc7038c545096747440cafea33cfc94f1f8d1", "precision": "fp32", "verification": "primary", "sources": [ "s3a9951ba0d608b", "s1c8a067400e270" ] }, { "age": 100.0, "six": 47.7025, "five": 50.71, "tasks": { "blimp_filtered": 69.77, "supplement_filtered": 59.78, "ewok_filtered": 51.3, "entity_tracking": 20.27, "comps": 52.43, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 43.0 }, "checkpoint": "ae2ddb00625aea4cc255a75b5b03a55e678ca882207b7d80eeb3b2830f147dde", "precision": "fp32", "verification": "primary", "sources": [ "se2d9dd0f3ca407", "sc217d4fec3c943" ] } ], "peak": { "five": { "score": 50.912, "ages": [ 80.0 ], "rank": 56 }, "six": { "score": 48.545, "ages": [ 80.0 ], "rank": 73 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 58, "id": "run/IPRLM384_GLOBAL_LR_RECIPE_SEED20260907_80M_20261023", "original_name": "IPRLM384_GLOBAL_LR_RECIPE_SEED20260907_80M_20261023", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "全体LR配合：本体500／Router520／Reader340-820-400／MoS500-450", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005, 0.0004375, 0.00037499999999999995 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40/70/80M after80M training" ], "intermediate": "saved40/70M states evaluated after80M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000, 70000000, 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000, 70000000, 80000000 ], "timing": "after80M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All31LRconditions plusparent320,40/70/80M and full historical inventory;same-age and peaks separated" }, "milestones_words": [ 40000000, 70000000, 80000000 ], "checkpoints_words": [ 40000000, 70000000, 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 108, "sources": [ "s5b118567b0bc74", "sb5ce1f2c409c39" ], "points": [ { "age": 40.0, "six": 47.58416666666667, "five": 50.074, "tasks": { "blimp_filtered": 69.07, "supplement_filtered": 57.62, "ewok_filtered": 52.0, "entity_tracking": 19.58, "comps": 52.1, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 46.0 }, "checkpoint": "24f8d443def3945ef09859caec20d9fd7d30c5ff495b917667725631cfe903ca", "precision": "fp32", "verification": "primary", "sources": [ "s8be3cbd40be3a4", "s2463259468df2c" ] }, { "age": 70.0, "six": 47.8075, "five": 50.73, "tasks": { "blimp_filtered": 69.95, "supplement_filtered": 59.39, "ewok_filtered": 52.11, "entity_tracking": 19.99, "comps": 52.21, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "00e8379cbf172b975dec9f73cf311cb4a2e4b1ee3ec5951b597f42fd550b92ec", "precision": "fp32", "verification": "primary", "sources": [ "s9dc90d94eab0c5", "s2463259468df2c" ] }, { "age": 80.0, "six": 48.5275, "five": 51.288, "tasks": { "blimp_filtered": 69.23, "supplement_filtered": 62.35, "ewok_filtered": 51.79, "entity_tracking": 20.36, "comps": 52.71, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 51.0 }, "checkpoint": "53b99e2ed900a5e5cf99e15d711318a8d7d6c9eaa016f99c9f437a1330ead134", "precision": "fp32", "verification": "primary", "sources": [ "sbcb7814f706de3", "s2463259468df2c" ] } ], "peak": { "five": { "score": 51.288, "ages": [ 80.0 ], "rank": 21 }, "six": { "score": 48.5275, "ages": [ 80.0 ], "rank": 74 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 59, "id": "run/IPRLM384_COUPLED_LR_REPAIR_SEED20260907_80M_20261024", "original_name": "IPRLM384_COUPLED_LR_REPAIR_SEED20260907_80M_20261024", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00082", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40/70/80M after80M training" ], "intermediate": "saved40/70M states evaluated after80M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000, 70000000, 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000, 70000000, 80000000 ], "timing": "after80M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All32LRconditions plusparent320;40/70/80M and full historical inventory;same-age and observed peaks separated" }, "milestones_words": [ 40000000, 70000000, 80000000 ], "checkpoints_words": [ 40000000, 70000000, 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 104, "sources": [ "s592673ed29544a" ], "points": [ { "age": 40.0, "six": 47.6675, "five": 50.162, "tasks": { "blimp_filtered": 69.47, "supplement_filtered": 57.99, "ewok_filtered": 51.05, "entity_tracking": 20.03, "comps": 52.27, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 50.0 }, "checkpoint": "8510818cc0be55216549b0c0e8eb4f33e4cd11f5cf6880b8cc7414250f4f6014", "precision": "fp32", "verification": "primary", "sources": [ "sed2eb73210d852", "sb78f2f8378ab5b" ] }, { "age": 70.0, "six": 48.524166666666666, "five": 50.396, "tasks": { "blimp_filtered": 70.3, "supplement_filtered": 58.32, "ewok_filtered": 51.56, "entity_tracking": 20.09, "comps": 51.71, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 56.0 }, "checkpoint": "c9f3616481937e829d6b0538543d06bd1cdffc5158f7cb4614a82bd604afb8da", "precision": "fp32", "verification": "primary", "sources": [ "s272c8dc591fbab", "sb78f2f8378ab5b" ] }, { "age": 80.0, "six": 48.455, "five": 50.622, "tasks": { "blimp_filtered": 69.8, "supplement_filtered": 59.12, "ewok_filtered": 51.4, "entity_tracking": 20.88, "comps": 51.91, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 50.0 }, "checkpoint": "9499f26830e091d936754e5622f1035039d1694f3ca3177c0291b2531970ccbb", "precision": "fp32", "verification": "primary", "sources": [ "sd2f0cb0c4dd99e", "sb78f2f8378ab5b" ] } ], "peak": { "five": { "score": 50.622, "ages": [ 80.0 ], "rank": 94 }, "six": { "score": 48.524166666666666, "ages": [ 70.0 ], "rank": 75 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 60, "id": "run/IPRLM_PARENT320_BRANCH123_W320_FFN1173_SEED20260907_20261018", "original_name": "IPRLM_PARENT320_BRANCH123_W320_FFN1173_SEED20260907_20261018", "title": "320幅・FFN1173・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1173, "depth": 9, "parts": { "width": 320, "ffn": 1173, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.665625 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1173, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_additive_bypass": false, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at70/80/90/100M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Full known inventory plus matched-age parent320; observed peaks retained separately" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 152, "sources": [ "s14c90bc2388c88", "sc0eee99c118a4a", "s5197ad6c5ee365" ], "points": [ { "age": 70.0, "six": 47.83416666666667, "five": 50.374, "tasks": { "blimp_filtered": 70.22, "supplement_filtered": 58.0, "ewok_filtered": 51.97, "entity_tracking": 19.33, "comps": 52.35, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 46.0 }, "checkpoint": "bba443d84428f51263b6a06dd274f7193529b6c3db83b431f33a9c09705512a8", "precision": "fp32", "verification": "primary", "sources": [ "s354f32833f8516", "s94f912c25057fd", "scbeace5ce8997b", "s2f727c6a91cb22" ] }, { "age": 80.0, "six": 48.4975, "five": 50.77, "tasks": { "blimp_filtered": 69.66, "supplement_filtered": 61.11, "ewok_filtered": 51.93, "entity_tracking": 18.52, "comps": 52.63, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 50.0 }, "checkpoint": "073a370e3dd7b327af2e79b3488c30e5ab85d3227822081c69e649176aa347a5", "precision": "fp32", "verification": "primary", "sources": [ "se14c537b0a181a", "s89f0bfffb3ab6a", "scbeace5ce8997b", "s2f727c6a91cb22" ] }, { "age": 90.0, "six": 48.221666666666664, "five": 50.342, "tasks": { "blimp_filtered": 69.83, "supplement_filtered": 58.11, "ewok_filtered": 51.28, "entity_tracking": 19.54, "comps": 52.95, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 50.0 }, "checkpoint": "3925273e56f2ac656881e7a3a6e61b732b27cd093c63f364edd6c4ec5a533401", "precision": "fp32", "verification": "primary", "sources": [ "s6e03609ec29168", "s0837eb2e4e9f53", "scbeace5ce8997b", "s2f727c6a91cb22" ] }, { "age": 100.0, "six": 47.94083333333333, "five": 50.202, "tasks": { "blimp_filtered": 69.92, "supplement_filtered": 58.45, "ewok_filtered": 50.47, "entity_tracking": 19.35, "comps": 52.82, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 49.0 }, "checkpoint": "7fe5caf97abf64e48504c3ebea70a86dfbd94841fb5414ce5c7201f5235877fe", "precision": "fp32", "verification": "primary", "sources": [ "sbd176389fef227", "sf590d0de7eb4ca", "scbeace5ce8997b", "s2f727c6a91cb22" ] } ], "peak": { "five": { "score": 50.77, "ages": [ 80.0 ], "rank": 79 }, "six": { "score": 48.4975, "ages": [ 80.0 ], "rank": 78 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 61, "id": "baby/B_READER_DEPTH9_BATCH4096_BETA2_100M", "original_name": "旧256・9層・単一出力", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "未確認", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 16, "sources": [ "s386714fb7a0434", "s8dda0fa37e2a5c", "s1cfa7ea0fe0025", "s885cc8e3b76bbe", "se94a2911ab5a5b", "s707c1f12f83adf", "s4f4631eaf27ff0", "sbfb8e3fff69005", "s63de9bf576bcfe" ], "points": [ { "age": 80, "six": 48.49333333333333, "five": 50.95, "tasks": { "blimp_filtered": 67.04, "supplement_filtered": 61.51, "ewok_filtered": 52.49, "entity_tracking": 21.18, "comps": 52.53, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 53.0 }, "checkpoint": "56e4f48c1c5c5a84dc9c5e9973c1b4e448acfd7c7c1c1b3e228c210d41a452b6", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "se808122f2fd715", "s15681f045c4791", "s01d80688fa59ff", "s3d246f98259bdb", "seee5953cf86717", "s729f05bc023b6f", "s776b1e2291b0b0", "s858f2c14cabaaa", "s45f2b7e3519558", "s2f5ea96a2c777a", "sc433646cb6b42b" ] }, { "age": 100, "six": 47.63166666666667, "five": 50.522, "tasks": { "blimp_filtered": 68.86, "supplement_filtered": 58.3, "ewok_filtered": 52.24, "entity_tracking": 20.88, "comps": 52.33, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 45.0 }, "checkpoint": "38ff36d003c7b24124d68d3c68ea974658e7af74b3a7f2fd3bc84cfc007a552d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "se808122f2fd715", "s15681f045c4791", "s1b68c8814d85ea", "sf4da6e6d20c6ac", "s729f05bc023b6f", "s776b1e2291b0b0", "s858f2c14cabaaa", "s45f2b7e3519558", "s2f5ea96a2c777a", "sc433646cb6b42b" ] } ], "peak": { "five": { "score": 50.95, "ages": [ 80 ], "rank": 50 }, "six": { "score": 48.49333333333333, "ages": [ 80 ], "rank": 79 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 62, "id": "run/IPRLM384_QK_RECEIVED_LR105_SEED20260907_40M_20261022", "original_name": "IPRLM384_QK_RECEIVED_LR105_SEED20260907_40M_20261022", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "QK・受信ゲートLR＋5％（共同LR版ベース）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.000504, 0.000441, 0.00037799999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M after40M training" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000, 70000000, 80000000 ], "timing": "after80M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All25LRconditions and fullmodelinventory;matchedRouter125parent" }, "milestones_words": [ 40000000, 70000000, 80000000 ], "checkpoints_words": [ 40000000, 70000000, 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 140, "sources": [ "s2aae068ce08fcf", "se088a71185eb44" ], "points": [ { "age": 40.0, "six": 48.20916666666667, "five": 50.63, "tasks": { "blimp_filtered": 69.18, "supplement_filtered": 60.0, "ewok_filtered": 52.05, "entity_tracking": 19.76, "comps": 52.16, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 46.0 }, "checkpoint": "75ca82e94abbdbd16aaa12a20c4db451efe4dc864d3054728aec7508b5762412", "precision": "fp32", "verification": "primary", "sources": [ "sb9c6a6587074e3", "s0c15fc15f0c2d6" ] }, { "age": 70.0, "six": 47.6375, "five": 50.226, "tasks": { "blimp_filtered": 69.37, "supplement_filtered": 57.95, "ewok_filtered": 50.88, "entity_tracking": 20.53, "comps": 52.4, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 49.0 }, "checkpoint": "01c1fc2bfd428f06d085674d724a54e72f4e0bb9e99f25c7f9c9d1cbbeffada7", "precision": "fp32", "verification": "primary", "sources": [ "sdae020d3dd89bc", "s0c15fc15f0c2d6" ] }, { "age": 80.0, "six": 48.47083333333333, "five": 50.832, "tasks": { "blimp_filtered": 68.99, "supplement_filtered": 61.24, "ewok_filtered": 51.85, "entity_tracking": 19.78, "comps": 52.3, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 51.0 }, "checkpoint": "d8445d67ece474181fab8408bdb268b5fa32c55b3a3e835efcd41ed03207ed6a", "precision": "fp32", "verification": "primary", "sources": [ "s72a047e4d167a5", "s0c15fc15f0c2d6" ] } ], "peak": { "five": { "score": 50.832, "ages": [ 80.0 ], "rank": 70 }, "six": { "score": 48.47083333333333, "ages": [ 80.0 ], "rank": 80 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 63, "id": "baby/vistarc_q3_mos3_delta_film320_20261010", "original_name": "320・MoS3・差分FiLM64", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_delta_film", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "output_relation_width": 64, "output_modulation_init": "zero", "output_delta_order": [ "h9-h8", "h8-h7" ], "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Matched50/60/70/80/90/100M and observed peaks versus320 nested parent and other frozen cohorts" }, "milestones_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 45, "sources": [ "s75ea05c09066ec", "sf846af9fdef03f", "sc7e7c39f59bace", "s0d71b9652db0cd", "sfcc8544ddfc8c9", "s8b28bcc61036a0", "s90f757df94d8cb" ], "points": [ { "age": 50, "six": 47.67, "five": 50.754, "tasks": { "blimp_filtered": 69.84, "supplement_filtered": 58.72, "ewok_filtered": 50.68, "entity_tracking": 21.67, "comps": 52.86, "global_piqa_parallel": 16.5, "global_piqa_nonparallel": 48.0 }, "checkpoint": "f3ed8e927f72a9fa7cbf3bf800467299ce479070bb735f756b6362af76779cd5", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s0241f51d0d6aa2", "s25d4fda5b8f793", "s2d0397e4c77a86", "sf5da492bab2fe4", "sb810f7de901b97", "sedc428601b6fdf", "s3444ba848d7474", "s82c51724f44c31" ] }, { "age": 60, "six": 48.458333333333336, "five": 50.62, "tasks": { "blimp_filtered": 69.98, "supplement_filtered": 58.33, "ewok_filtered": 52.11, "entity_tracking": 19.98, "comps": 52.7, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 52.0 }, "checkpoint": "ec682f95e243f92f4cbddbca7688d509e8a50b6c737b7a601af177739716c6cf", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sacadfc014a876c", "s18e2d2341a596c", "s7d5d80e76274ea", "sf5da492bab2fe4", "sb810f7de901b97", "sedc428601b6fdf", "s3444ba848d7474", "s82c51724f44c31" ] }, { "age": 70, "six": 48.33083333333333, "five": 50.776, "tasks": { "blimp_filtered": 70.67, "supplement_filtered": 59.03, "ewok_filtered": 52.26, "entity_tracking": 19.63, "comps": 52.29, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 46.0 }, "checkpoint": "aecbf205dc6873944457eb8b1d23b076a9c963ad361a62a0a690b50456fb0d88", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s81ad739660f643", "sdb8acae57b989a", "s16a6afc253c27c", "sf5da492bab2fe4", "sb810f7de901b97", "sedc428601b6fdf", "s3444ba848d7474", "s82c51724f44c31" ] }, { "age": 80, "six": 48.1025, "five": 50.678, "tasks": { "blimp_filtered": 68.74, "supplement_filtered": 60.16, "ewok_filtered": 52.08, "entity_tracking": 20.0, "comps": 52.41, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 52.0 }, "checkpoint": "53ccc343ecc0b2129acd57c6eac9eda26ee09ca339f866976d7a7ed6f55fd86d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s482e0f5059be5c", "s3a9b6aa09f559d", "sf64c047ac8924b", "sf5da492bab2fe4", "sb810f7de901b97", "sedc428601b6fdf", "s3444ba848d7474", "s82c51724f44c31" ] }, { "age": 90, "six": 48.16583333333333, "five": 50.666, "tasks": { "blimp_filtered": 68.94, "supplement_filtered": 58.8, "ewok_filtered": 53.42, "entity_tracking": 19.73, "comps": 52.44, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 49.0 }, "checkpoint": "099d7746df0a10d213c478329808b6b36d439fc04e125dccb0de2995426883fc", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s87b638755c5c7e", "s6e7b16552b05d5", "s34dcc6a65517e7", "sf5da492bab2fe4", "sb810f7de901b97", "sedc428601b6fdf", "s3444ba848d7474", "s82c51724f44c31" ] }, { "age": 100, "six": 48.1075, "five": 51.09, "tasks": { "blimp_filtered": 68.7, "supplement_filtered": 60.21, "ewok_filtered": 53.72, "entity_tracking": 20.23, "comps": 52.59, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "d4804f8d956b83b5e4a3c892a255d668db63d99c33408b2909a5db675a8b83a1", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s213666c1673799", "s9f80f267487f6f", "sc178ed644c7e24", "sf5da492bab2fe4", "sb810f7de901b97", "sedc428601b6fdf", "s3444ba848d7474", "s82c51724f44c31" ] } ], "peak": { "five": { "score": 51.09, "ages": [ 100 ], "rank": 36 }, "six": { "score": 48.458333333333336, "ages": [ 60 ], "rank": 82 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 64, "id": "run/IPRLM320_FFN_CONDITIONAL_BYPASS_SEED137_20261015", "original_name": "IPRLM320_FFN_CONDITIONAL_BYPASS_SEED137_20261015", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "negative_sigmoid", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true, "ffn_bypass_condition": "negative_sigmoid", "ffn_bypass_threshold": 5.0 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After each40M training; no automatic extension.", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed same40M parent320, unconditional bypass and conditional bypass. 40M is budget allocation, not final peak ranking." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 78, "sources": [ "sfb0e47107f86f4", "s7ff79e5101a18a", "s173cba363a4725" ], "points": [ { "age": 40.0, "six": 47.54833333333333, "five": 50.516, "tasks": { "blimp_filtered": 68.73, "supplement_filtered": 58.51, "ewok_filtered": 52.25, "entity_tracking": 21.41, "comps": 51.68, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 46.0 }, "checkpoint": "2ce044b76829db1d3e10c9d87562466a46d2652f1a5f80a62c3b3ac20034c819", "precision": "fp32", "verification": "primary", "sources": [ "se210f9d87f1a25", "s812e0c07737aa4" ] }, { "age": 70.0, "six": 48.439166666666665, "five": 50.888, "tasks": { "blimp_filtered": 69.6, "supplement_filtered": 59.4, "ewok_filtered": 53.11, "entity_tracking": 19.87, "comps": 52.46, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 52.0 }, "checkpoint": "44b15245d816294ca501363f677de3128d1507dd91e143b422f251193a3a3217", "precision": "fp32", "verification": "primary", "sources": [ "seaf807e2ac0567", "s7a80a0980ae060" ] } ], "peak": { "five": { "score": 50.888, "ages": [ 70.0 ], "rank": 61 }, "six": { "score": 48.439166666666665, "ages": [ 70.0 ], "rank": 83 } }, "clear_points": [ { "number": 64, "age": 70, "seed": 137, "checkpoint_sha256": "44b15245d816294ca501363f677de3128d1507dd91e143b422f251193a3a3217", "old_six": 48.439166666666665, "old_five": 50.888, "old_tasks": { "blimp_filtered": 69.6, "supplement_filtered": 59.4, "ewok_filtered": 53.11, "entity_tracking": 19.87, "comps": 52.46, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 52.0 }, "correct": 12244, "total": 17274, "accuracy": 70.8810929720968, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6962, "total": 8392, "ties": 0 }, "comps": { "correct": 3923, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 1, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 20, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 956, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 72, "total": 114, "ties": 0 }, "boolq": { "correct": 9, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 66.01656670300144, "checkpoint_rank": 11, "rank": 10 } ], "clear_peak": { "number": 64, "age": 70, "seed": 137, "checkpoint_sha256": "44b15245d816294ca501363f677de3128d1507dd91e143b422f251193a3a3217", "old_six": 48.439166666666665, "old_five": 50.888, "old_tasks": { "blimp_filtered": 69.6, "supplement_filtered": 59.4, "ewok_filtered": 53.11, "entity_tracking": 19.87, "comps": 52.46, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 52.0 }, "correct": 12244, "total": 17274, "accuracy": 70.8810929720968, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6962, "total": 8392, "ties": 0 }, "comps": { "correct": 3923, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 1, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 20, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 956, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 72, "total": 114, "ties": 0 }, "boolq": { "correct": 9, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 66.01656670300144, "checkpoint_rank": 11, "rank": 10 }, "present": true }, { "number": 65, "id": "run/Q3_MOS3_BRANCH123_WIDTH384_MATCHED320_UPDATES_20261011", "original_name": "Q3_MOS3_BRANCH123_WIDTH384_MATCHED320_UPDATES_20261011", "title": "384幅・FFN1344・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1344, "depth": 9, "parts": { "width": 384, "ffn": 1344, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1344, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_init_only", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes70/80/90/100M", "SyntaxGym31+PIQA1838 at70/80/90/100M", "frozen bAbI1947+ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 TF32off exact prefix graph; actual384 speed, numerical parity and all capacity buckets verified", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Matched ages and observed peaks separately, including ordinary320 shared-last2 and all selected historical models" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ], "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260, "driver_compatibility": { "scope": "system libraries after reboot; no LD_LIBRARY_PATH override", "kernel_version": "595.99.02", "library_version": "595.99.02", "system_driver_modified_by_agent": false, "post_reboot_proof": "POST_REBOOT_ACCEPTANCE.json" } }, "condition_conflicts": [], "profile": 146, "sources": [ "s7130c333e95db1", "s2bb1e08f3d7e9e", "sf30ead3aeb3b91", "s9392761bd7fed5", "scb5dd622d0f3cb", "s712b36a604f445", "s3af294dc4e6fae", "sab9a038c27ef75" ], "points": [ { "age": 40.0, "six": 48.0325, "five": 50.206, "tasks": { "blimp_filtered": 68.39, "supplement_filtered": 59.84, "ewok_filtered": 51.25, "entity_tracking": 19.32, "comps": 52.23, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "checkpoint": "03174aac15fe625581dec27f3a1886f069a6d4d985e1b24aaed22c614bd86761", "precision": "fp32", "verification": "primary", "sources": [ "s51f10b78ab8dfe", "s96e9a0d79c27d1" ] }, { "age": 70.0, "six": 48.24666666666667, "five": 50.754, "tasks": { "blimp_filtered": 69.8, "supplement_filtered": 59.26, "ewok_filtered": 52.2, "entity_tracking": 20.2, "comps": 52.31, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 52.0 }, "checkpoint": "2f9aa9f9810afc7690e479ed8c5e4501f34e8675cc5a97a49a8d493ba0ba3357", "precision": "fp32", "verification": "primary", "sources": [ "sb84fd06afefd8b", "s852f5126ec972a", "seedd7108b1a63f", "s6b8d5b59996463", "s827d5168aebefe", "s5ae8d4fa270b37", "s2a7d8025ec9454", "s76c5fb4fbeca40" ] }, { "age": 80.0, "six": 48.43333333333333, "five": 50.978, "tasks": { "blimp_filtered": 68.0, "supplement_filtered": 61.84, "ewok_filtered": 52.23, "entity_tracking": 20.77, "comps": 52.05, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 52.0 }, "checkpoint": "4c594f50ed315e7f58207e7efa5f66ed6fad50c2d562056d472ad9eb815a1d1c", "precision": "fp32", "verification": "primary", "sources": [ "scebb78fc47143d", "s522eddffff577a", "s7c5212da7a1622", "s6b8d5b59996463", "s827d5168aebefe", "s5ae8d4fa270b37", "s2a7d8025ec9454", "s76c5fb4fbeca40" ] }, { "age": 90.0, "six": 46.975, "five": 49.734, "tasks": { "blimp_filtered": 68.77, "supplement_filtered": 57.25, "ewok_filtered": 50.51, "entity_tracking": 19.96, "comps": 52.18, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 45.0 }, "checkpoint": "9408e21dae9fc0ea6e2138baaf1543345bf973fe4f9f3e564e4fc803ee0ee96f", "precision": "fp32", "verification": "primary", "sources": [ "sfb8f08f8e66481", "sc5eb782f76bb66", "s15761c4d26155d", "s6b8d5b59996463", "s827d5168aebefe", "s5ae8d4fa270b37", "s2a7d8025ec9454", "s76c5fb4fbeca40" ] }, { "age": 100.0, "six": 47.71333333333333, "five": 50.62, "tasks": { "blimp_filtered": 68.1, "supplement_filtered": 62.31, "ewok_filtered": 50.62, "entity_tracking": 19.65, "comps": 52.42, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 45.0 }, "checkpoint": "443f60938b424908c897e527ac7a175d96a0e55fe4d278c9ae895471719a9a20", "precision": "fp32", "verification": "primary", "sources": [ "s5e95f44390c367", "s47c2eb0744d2db", "sd3296f8a860a86", "s6b8d5b59996463", "s827d5168aebefe", "s5ae8d4fa270b37", "s2a7d8025ec9454", "s76c5fb4fbeca40" ] } ], "peak": { "five": { "score": 50.978, "ages": [ 80.0 ], "rank": 48 }, "six": { "score": 48.43333333333333, "ages": [ 80.0 ], "rank": 84 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 66, "id": "baby/vistarc_q3_mos3_last5_received_value_gate_20261006", "original_name": "256・MoS3・5層混合", "title": "256幅・FFN896・9層", "topology": "MoS3・共通5層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "5・6・7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 5, 6, 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last5", "attention_received_gate": true, "output_components": 3, "output_state_layers": 5, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3_LAST5 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "metric": "maximum observed local six-axis", "train_words": 100000000, "save_words": [ 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M training completes; serial", "historical_record": 49.315, "no_endpoint_penalty": true, "diagnostics": [ "same-checkpoint five-axis and seven component scores", "training loss", "candidate and reader usage", "mixture usage", "runtime" ], "comparison_scope": "Same-cohort2/3/4/5 source states,9L MoS3,70/80/90/100M" } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 54, "sources": [ "s9acb97f35e573b", "s22eb23b365d1ed", "sc54f84281c2c22", "sef7b3714696d8a", "s7de53076e02084", "s3aa117519903e1", "sc7265384b499ad" ], "points": [ { "age": 70, "six": 48.153333333333336, "five": 50.848, "tasks": { "blimp_filtered": 70.1, "supplement_filtered": 59.16, "ewok_filtered": 50.46, "entity_tracking": 22.06, "comps": 52.46, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 48.0 }, "checkpoint": "dc5a81100036ade3334d76ff75515e2a63df142bdb44c8c57b5929484f6eca3d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5717258d8cb382", "sbdff20de9df691", "se906613eadf114", "sebdf2a03cdd8f2", "sa7a4070dacae4f", "s775c47d17ba3dd" ] }, { "age": 80, "six": 48.38916666666667, "five": 50.534, "tasks": { "blimp_filtered": 68.69, "supplement_filtered": 61.15, "ewok_filtered": 51.47, "entity_tracking": 19.39, "comps": 51.97, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 53.0 }, "checkpoint": "d8b39997bd4212c712b795ac68e070b012f838ce475c28f78355d59e5d507b71", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1ade0b4f684746", "sbdff20de9df691", "se906613eadf114", "sebdf2a03cdd8f2", "sa7a4070dacae4f", "s775c47d17ba3dd" ] }, { "age": 90, "six": 46.88666666666666, "five": 49.928, "tasks": { "blimp_filtered": 68.52, "supplement_filtered": 58.6, "ewok_filtered": 52.38, "entity_tracking": 18.24, "comps": 51.9, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 42.0 }, "checkpoint": "6bb85e767ecdce3f9216b4b10404bcf53d1959f49b7e55611a491072380305b4", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s00fd810f9cff7f", "sbdff20de9df691", "se906613eadf114", "sebdf2a03cdd8f2", "sa7a4070dacae4f", "s775c47d17ba3dd" ] }, { "age": 100, "six": 47.288333333333334, "five": 50.11, "tasks": { "blimp_filtered": 70.34, "supplement_filtered": 57.77, "ewok_filtered": 51.14, "entity_tracking": 19.08, "comps": 52.22, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 45.0 }, "checkpoint": "67af476926edb6e18107b707039036c1fe9105ea32226e97133994b6101423a7", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sde1ae2d45475cc", "sbdff20de9df691", "se906613eadf114", "sebdf2a03cdd8f2", "sa7a4070dacae4f", "s775c47d17ba3dd" ] } ], "peak": { "five": { "score": 50.848, "ages": [ 70 ], "rank": 68 }, "six": { "score": 48.38916666666667, "ages": [ 80 ], "rank": 85 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 67, "id": "baby/vistarc_q3_mos3_last2_width320_ffn1120_heads10_20261009", "original_name": "320・MoS3・通常2層混合", "title": "320幅・FFN1120・9層", "topology": "MoS3・共通2層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last2", "attention_received_gate": true, "output_components": 3, "output_state_layers": 2, "reference_attention_width": 256, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3_LAST2 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.715, "comparison_scope": "Matched80M and observed peaks separately; 256/272/384 MoS3 last2 and other MoS baselines" }, "milestones_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 51, "sources": [ "sbfbea759bdc9dd", "s26dab7c709e3e1", "s14ef5013292c49", "sbd1d9361761cbd", "sf669c98d2b2c4e", "s171ac733164fe9", "s4b25e5b2ef54cf", "s1347042f1da6f8", "s2be536f80a1ff0", "s3ec5b1e1ff2eda" ], "points": [ { "age": 40.0, "six": 48.376666666666665, "five": 50.322, "tasks": { "blimp_filtered": 69.26, "supplement_filtered": 56.31, "ewok_filtered": 52.56, "entity_tracking": 21.11, "comps": 52.37, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 54.0 }, "checkpoint": "e966646954dffc62ee7b4392050dc454095b8c9d5f950dd52078ecfd10705750", "precision": "fp32", "verification": "primary", "sources": [ "s18b3473c505ab7", "sd5d472e531910d" ] }, { "age": 50, "six": 47.42, "five": 50.362, "tasks": { "blimp_filtered": 69.41, "supplement_filtered": 57.59, "ewok_filtered": 51.4, "entity_tracking": 21.07, "comps": 52.34, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 46.0 }, "checkpoint": "8addf122987bbe2534992eabbefad9b24bbc34c8be50b38e2200c0c387fbea77", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "seb99902aaa367b", "sa2d26b73b6390b", "s81d335e92a9f43", "s111dcaa8f41c87", "sdb6b71b1a7adcc", "s96188d4eef490b", "s85c8cfe36f2adb", "sb58b7f967c1965" ] }, { "age": 60, "six": 47.895833333333336, "five": 49.842, "tasks": { "blimp_filtered": 68.24, "supplement_filtered": 56.36, "ewok_filtered": 52.78, "entity_tracking": 19.97, "comps": 51.86, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 54.0 }, "checkpoint": "35a668a5f87c56383c4d25cc2c0ba5f9442d64c9fa98652723a9159e7b1bcbbd", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s38bff2ea6337e8", "s4f6fc738eb7e02", "s18344847b169a1", "s111dcaa8f41c87", "sdb6b71b1a7adcc", "s96188d4eef490b", "s85c8cfe36f2adb", "sb58b7f967c1965" ] }, { "age": 70, "six": 47.58, "five": 50.066, "tasks": { "blimp_filtered": 69.0, "supplement_filtered": 57.36, "ewok_filtered": 51.16, "entity_tracking": 20.39, "comps": 52.42, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 47.0 }, "checkpoint": "3d52b89d1919b17b962b90e6b057e3fe3dda7464931873cd2df3f8635d8e2743", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sd36674a508a53f", "s0fef30a97757c9", "secbaa2d4e92d7b", "s111dcaa8f41c87", "sdb6b71b1a7adcc", "s96188d4eef490b", "s85c8cfe36f2adb", "sb58b7f967c1965" ] }, { "age": 80, "six": 48.32, "five": 50.46, "tasks": { "blimp_filtered": 68.68, "supplement_filtered": 60.08, "ewok_filtered": 50.81, "entity_tracking": 20.47, "comps": 52.26, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 50.0 }, "checkpoint": "e3bbb79d4dd0996e69d8dfe16bbd91ca1bfbea77727cb006278343a13f276055", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s98881c97911e3e", "se0ef79250c1659", "s4d2b91ebff0bf2", "s111dcaa8f41c87", "sdb6b71b1a7adcc", "s96188d4eef490b", "s85c8cfe36f2adb", "sb58b7f967c1965" ] }, { "age": 90, "six": 47.56666666666667, "five": 50.344, "tasks": { "blimp_filtered": 69.11, "supplement_filtered": 57.0, "ewok_filtered": 52.1, "entity_tracking": 21.03, "comps": 52.48, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 46.0 }, "checkpoint": "c67af0b51a31f30644a92ece74fd37bf149f68ebfade0e86a6c4762387b058e7", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s9043364c040ab1", "sc61fc05cd17598", "s1ba7bf70c5750b", "s111dcaa8f41c87", "sdb6b71b1a7adcc", "s96188d4eef490b", "s85c8cfe36f2adb", "sb58b7f967c1965" ] }, { "age": 100, "six": 47.58, "five": 50.466, "tasks": { "blimp_filtered": 69.22, "supplement_filtered": 58.47, "ewok_filtered": 50.98, "entity_tracking": 21.3, "comps": 52.36, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 43.0 }, "checkpoint": "a681f0e3c06d2070c083c2aefd269550101d1e7364583aca81b3dde365993eef", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s9ce8479626a218", "s5666663eb0405d", "s4fca7967a890e3", "s111dcaa8f41c87", "sdb6b71b1a7adcc", "s96188d4eef490b", "s85c8cfe36f2adb", "sb58b7f967c1965" ] } ], "peak": { "five": { "score": 50.466, "ages": [ 100 ], "rank": 114 }, "six": { "score": 48.376666666666665, "ages": [ 40.0 ], "rank": 87 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 68, "id": "baby/vistarc_q3_single_output_received_gate_20261005", "original_name": "256・単一出力＋受信ゲート", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single_linear", "attention_received_gate": true, "ffn_input_output_gates": false, "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ], "objective": { "metric": "maximum observed local six-axis", "train_words": 100000000, "save_words": [ 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "benchmark_words": [ 80000000, 100000000 ], "timing": "after100M training completes; serial", "parent_peak": 48.75666666666667, "historical_record": 49.03583333333333, "record_gap_from_parent": 0.27916666666666146, "no_endpoint_penalty": true, "diagnostics": [ "same-checkpoint five-axis and seven component scores", "training loss", "candidate and reader usage", "runtime" ], "comparison_scope": "Same80/100M checkpoints of Q3 MoS2 received-gate parent versus removal of MoS2 only" } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 64, "sources": [ "s61b55eeb904358", "sc53cd3875595fb", "se4fc79015844a3", "sbe40ae4306789d", "sb0a4695f05c3e5", "s92625421b8654b", "sd2ae6532b17709", "s661f1a32b1fd60", "s61fc2804cede4a" ], "points": [ { "age": 70, "six": 48.22083333333333, "five": 50.538, "tasks": { "blimp_filtered": 68.61, "supplement_filtered": 59.2, "ewok_filtered": 51.87, "entity_tracking": 20.36, "comps": 52.65, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 49.0 }, "checkpoint": "b979f150378f2a1ee03811a8dec630eb0162547861ef6b9e65bd7ec24d64a4cc", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s424074126234bd", "s57df6a5c3efa09", "s3db0dc2cdd1900", "s2489b8112cef84", "s8336e965d26a2f", "s3b4aba87343567", "sbc8934b76d1e87", "s85a3de752e083b", "scc8a44029b983e" ] }, { "age": 80, "six": 48.045, "five": 50.424, "tasks": { "blimp_filtered": 67.87, "supplement_filtered": 59.81, "ewok_filtered": 50.81, "entity_tracking": 21.03, "comps": 52.6, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 49.0 }, "checkpoint": "0df91ad57055bfe5df4563e30e0a1ccc7f461a1525366979263f4fbb23a022ac", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "seb1ae36f696101", "s57df6a5c3efa09", "s3db0dc2cdd1900", "s9f61dcfc259510", "s8336e965d26a2f", "s3b4aba87343567", "sbc8934b76d1e87", "s85a3de752e083b", "scc8a44029b983e", "s97e52f5d709ea0" ] }, { "age": 90, "six": 47.015, "five": 49.488, "tasks": { "blimp_filtered": 67.24, "supplement_filtered": 57.48, "ewok_filtered": 49.91, "entity_tracking": 20.28, "comps": 52.53, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 46.0 }, "checkpoint": "363fabc772d5ff6c92b19c1bc8bf3e7364b4e55d0c01f2de3537844e17b3bb56", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sfb5a7ed7925507", "s57df6a5c3efa09", "s3db0dc2cdd1900", "s348543adb35b1e", "s8336e965d26a2f", "s3b4aba87343567", "sbc8934b76d1e87", "s85a3de752e083b", "scc8a44029b983e" ] }, { "age": 100, "six": 48.376666666666665, "five": 50.334, "tasks": { "blimp_filtered": 67.57, "supplement_filtered": 59.56, "ewok_filtered": 51.58, "entity_tracking": 20.36, "comps": 52.6, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 50.0 }, "checkpoint": "a5db45d345c26f9bb1276e2fb0795abef50a4077a4a0d94ad2e4ef1397738acf", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc01d1b18c6e0f0", "s57df6a5c3efa09", "s3db0dc2cdd1900", "s6542e9ebfd00e3", "s8336e965d26a2f", "s3b4aba87343567", "sbc8934b76d1e87", "s85a3de752e083b", "scc8a44029b983e", "s97e52f5d709ea0" ] } ], "peak": { "five": { "score": 50.538, "ages": [ 70 ], "rank": 107 }, "six": { "score": 48.376666666666665, "ages": [ 100 ], "rank": 87 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 69, "id": "baby/B_Q0_MULTI_INPUT_MOS2_100M_20261004", "original_name": "Q0・MoS2・3層参照", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos2_last3", "attention_received_gate": false, "ffn_input_output_gates": false, "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260, "fused_mixture": "eager-equivalent Triton FP32 two-term logsumexp;native broadcast reduction backward" }, "condition_conflicts": [], "profile": 9, "sources": [ "s696611a6457bd4", "s6e1c8066392b49", "scce5772713e00e", "s6c13b922baba22", "s41129f04f6cc18", "s86f28d323d9f14" ], "points": [ { "age": 80, "six": 48.373333333333335, "five": 50.124, "tasks": { "blimp_filtered": 67.88, "supplement_filtered": 60.4, "ewok_filtered": 50.01, "entity_tracking": 20.58, "comps": 51.75, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 54.0 }, "checkpoint": "74a5bba1cd9ddd9d387a62d4a310b0127fc0b861dbbcdb535fd677f6f9ab576f", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1507d53d8621bb", "s494333ca4cd4c1", "s27623d0cd3357d", "s4bde78e342629b", "s54da4dc923a4d4", "s3e9c0b527e045c", "scb92ac3cfeb317", "sdcc3085efce423" ] }, { "age": 100, "six": 47.4575, "five": 50.01, "tasks": { "blimp_filtered": 68.77, "supplement_filtered": 58.5, "ewok_filtered": 50.82, "entity_tracking": 19.79, "comps": 52.17, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 49.0 }, "checkpoint": "98f0f66a3d578066a07d90dbe4acc5172b2d10f5aec761f0d1f48da742956b65", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1507d53d8621bb", "s5f2387c6e3ff2d", "s27623d0cd3357d", "s88488477753fe3", "s54da4dc923a4d4", "s3e9c0b527e045c", "scb92ac3cfeb317", "sdcc3085efce423" ] } ], "peak": { "five": { "score": 50.124, "ages": [ 80 ], "rank": 142 }, "six": { "score": 48.373333333333335, "ages": [ 80 ], "rank": 89 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 70, "id": "run/IPRLM320_FFN_CONDITIONAL_BYPASS_SEED20260907_20261015", "original_name": "IPRLM320_FFN_CONDITIONAL_BYPASS_SEED20260907_20261015", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "negative_sigmoid", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true, "ffn_bypass_condition": "negative_sigmoid", "ffn_bypass_threshold": 5.0 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After each40M training; no automatic extension.", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed same40M parent320, unconditional bypass and conditional bypass. 40M is budget allocation, not final peak ranking." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 79, "sources": [ "saf2855f8a98e20", "s63d65908786576", "s3b2e39e23ac6d8" ], "points": [ { "age": 40.0, "six": 47.8775, "five": 50.22, "tasks": { "blimp_filtered": 69.12, "supplement_filtered": 57.21, "ewok_filtered": 52.08, "entity_tracking": 20.49, "comps": 52.2, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 50.0 }, "checkpoint": "08104d60fbddb28dbf9b41460a0330f4e0f7ffe211d9c63e1352091d46e5f055", "precision": "fp32", "verification": "primary", "sources": [ "s400309195182af", "see0b7f448458db" ] }, { "age": 70.0, "six": 48.36416666666667, "five": 50.604, "tasks": { "blimp_filtered": 69.31, "supplement_filtered": 57.68, "ewok_filtered": 51.6, "entity_tracking": 22.25, "comps": 52.18, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "checkpoint": "d97f968ca9eb84efe2a89df5a9c3532c1bcd7b9c0bf888e2b030566873ce3736", "precision": "fp32", "verification": "primary", "sources": [ "s6395a68308a15f", "s2406f10a5f781b" ] } ], "peak": { "five": { "score": 50.604, "ages": [ 70.0 ], "rank": 97 }, "six": { "score": 48.36416666666667, "ages": [ 70.0 ], "rank": 90 } }, "clear_points": [ { "number": 70, "age": 70, "seed": 20260907, "checkpoint_sha256": "d97f968ca9eb84efe2a89df5a9c3532c1bcd7b9c0bf888e2b030566873ce3736", "old_six": 48.36416666666667, "old_five": 50.604, "old_tasks": { "blimp_filtered": 69.31, "supplement_filtered": 57.68, "ewok_filtered": 51.6, "entity_tracking": 22.25, "comps": 52.18, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "correct": 12196, "total": 17274, "accuracy": 70.6032187102003, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6924, "total": 8392, "ties": 0 }, "comps": { "correct": 3887, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 4, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 12, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 985, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 73, "total": 114, "ties": 0 }, "boolq": { "correct": 9, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 212, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 65.86184164800216, "checkpoint_rank": 13, "rank": 12 } ], "clear_peak": { "number": 70, "age": 70, "seed": 20260907, "checkpoint_sha256": "d97f968ca9eb84efe2a89df5a9c3532c1bcd7b9c0bf888e2b030566873ce3736", "old_six": 48.36416666666667, "old_five": 50.604, "old_tasks": { "blimp_filtered": 69.31, "supplement_filtered": 57.68, "ewok_filtered": 51.6, "entity_tracking": 22.25, "comps": 52.18, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "correct": 12196, "total": 17274, "accuracy": 70.6032187102003, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6924, "total": 8392, "ties": 0 }, "comps": { "correct": 3887, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 4, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 12, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 985, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 73, "total": 114, "ties": 0 }, "boolq": { "correct": 9, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 212, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 65.86184164800216, "checkpoint_rank": 13, "rank": 12 }, "present": true }, { "number": 71, "id": "baby/Q5_READER48_RESPONSIVE", "original_name": "Q5・Reader48 Responsive", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 33, "sources": [ "sa7c3403cfe3906", "s61371e9ed44fd1", "s9f8c27662df5ae", "s37bf318829e935", "s5c262b16e87429" ], "points": [ { "age": 80, "six": 48.344166666666666, "five": 50.886, "tasks": { "blimp_filtered": 68.88, "supplement_filtered": 59.52, "ewok_filtered": 53.06, "entity_tracking": 20.56, "comps": 52.41, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 47.0 }, "checkpoint": "5b449d40b54c10e4c80789d569f3d085c7814d8e1ff135a765e54a89f2d02a96", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s94b26eed584df4", "s738997fa713d8c", "s77883af07bb9ed", "s9d3c8af2b31c4d", "s7f5e022e28aef4" ] }, { "age": 100, "six": 47.51416666666667, "five": 50.184, "tasks": { "blimp_filtered": 68.46, "supplement_filtered": 58.02, "ewok_filtered": 51.22, "entity_tracking": 20.68, "comps": 52.54, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "4d865d7aad7bc9b1654e2bf59f2c19c7ed9fd11ac8414bd54fcbe6568fcb05f0", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s94b26eed584df4", "s738997fa713d8c", "s5f1bfb6567c660", "s9d3c8af2b31c4d", "s7f5e022e28aef4" ] } ], "peak": { "five": { "score": 50.886, "ages": [ 80 ], "rank": 62 }, "six": { "score": 48.344166666666666, "ages": [ 80 ], "rank": 92 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 72, "id": "baby/vistarc_q3_single_output_received_ffn_gates_20261005", "original_name": "256・単一出力＋受信・FFN両ゲート", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "あり", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single_linear", "attention_received_gate": true, "ffn_input_output_gates": true, "ffn_gate_rank": 32, "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ], "objective": { "metric": "maximum observed local six-axis", "train_words": 100000000, "save_words": [ 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "benchmark_words": [ 80000000, 100000000 ], "timing": "after100M training completes; serial", "historical_record": 49.03583333333333, "no_endpoint_penalty": true, "diagnostics": [ "same-checkpoint five-axis and seven component scores", "training loss", "candidate and reader usage", "runtime" ], "comparison_scope": "Same80/100M local six-axis comparison with single-output received-gate parent; peak objective" } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 63, "sources": [ "s8e84094bc3dfcd", "s8fdbcdd9f1a0fd", "s5f23c3f56964aa", "s504d4f884d77fd", "sd6abe6a1a3e400", "s4ed3dfdc2abafc", "seba51c61a19267", "s3adee0ffeb1d25" ], "points": [ { "age": 80, "six": 48.3425, "five": 50.772, "tasks": { "blimp_filtered": 69.33, "supplement_filtered": 60.4, "ewok_filtered": 52.15, "entity_tracking": 19.23, "comps": 52.75, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 52.0 }, "checkpoint": "b8c540a8af48e9252b96eb28c091f1151396ac9c47b651eb847823deacf101af", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s9693492015dec2", "s29fb05c5340b8d", "s24d22e40e40847", "sef6db0515f0b41", "s981ced5a9be2ff", "s6e0bb3492447dc", "sc16bbd37755015", "sf6e34bec953206" ] }, { "age": 100, "six": 48.17166666666667, "five": 50.782, "tasks": { "blimp_filtered": 70.01, "supplement_filtered": 58.56, "ewok_filtered": 52.89, "entity_tracking": 19.65, "comps": 52.8, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 45.0 }, "checkpoint": "cd79cacd361ff64b9c1b358673e1bdeaeaa68d71ffa1ebf3c5e820b79780c8b3", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s9cdb80031ae11a", "s29fb05c5340b8d", "s24d22e40e40847", "s8831d161e2dbcf", "s981ced5a9be2ff", "s6e0bb3492447dc", "sc16bbd37755015", "sf6e34bec953206" ] } ], "peak": { "five": { "score": 50.782, "ages": [ 100 ], "rank": 78 }, "six": { "score": 48.3425, "ages": [ 80 ], "rank": 94 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 73, "id": "baby/B_READER_DEPTH9_WIDTH272_ATTN256_READER64_LEARNED_100M", "original_name": "旧272・Attention256・Reader64可変", "title": "272幅・FFN960・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 272, "ffn": 960, "depth": 9, "parts": { "width": 272, "ffn": 960, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 64, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5294117647058822 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 272, "layers": 9, "hidden": 960, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 64, "history_norm": "separate", "seed": 20260907, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "attention_width": 256, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score60/70/80/90/100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 21, "sources": [ "sbc91420963cad6", "s81b88a6f712d7c", "sf725eeb23bec06", "s10feeb43476496", "s1b46af5c697d8d", "sc719ad9f741c28", "sd3d2f961bc0de8" ], "points": [ { "age": 60, "six": 47.31916666666667, "five": 49.662, "tasks": { "blimp_filtered": 67.05, "supplement_filtered": 56.65, "ewok_filtered": 53.06, "entity_tracking": 19.32, "comps": 52.23, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 45.0 }, "checkpoint": "6c4a52cfe16e66ecc0f2d6cbdccaac829a87d9b7c66146c015256d8203451840", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sfa13e100cd7783", "s35620960af7472", "s68f2556326de1a", "s32ebfb105f1400", "sceeff5dc385420", "sf70484a056af7f", "sfb751107480370", "s0d1c545a72fca0", "s844f17699103d6" ] }, { "age": 70, "six": 48.22, "five": 50.434, "tasks": { "blimp_filtered": 68.36, "supplement_filtered": 57.95, "ewok_filtered": 53.55, "entity_tracking": 19.68, "comps": 52.63, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "6403c60be38909d09a4d8751d82875e04bb9f3d2f40e756227ccbd85031fe757", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sfa13e100cd7783", "s35620960af7472", "s5be42c678fad34", "s5c05a3da0f2484", "sceeff5dc385420", "sf70484a056af7f", "sfb751107480370", "s0d1c545a72fca0", "s844f17699103d6" ] }, { "age": 80, "six": 48.32083333333333, "five": 50.552, "tasks": { "blimp_filtered": 67.9, "supplement_filtered": 62.02, "ewok_filtered": 51.95, "entity_tracking": 18.28, "comps": 52.61, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "checkpoint": "8ac96828039dde2465c39355dc366694158f1eed8286886813922ca8b72e4f9b", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sfa13e100cd7783", "s35620960af7472", "s8336127cf809dd", "s4f0ccc0c3f35cd", "sceeff5dc385420", "sf70484a056af7f", "sfb751107480370", "s0d1c545a72fca0", "s844f17699103d6" ] }, { "age": 90, "six": 47.9075, "five": 49.962, "tasks": { "blimp_filtered": 67.41, "supplement_filtered": 59.1, "ewok_filtered": 51.22, "entity_tracking": 19.63, "comps": 52.45, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 51.0 }, "checkpoint": "a0193ea1f5b07ca339ea85d053bf4adc531929c92142fae8a5ee41acbfc98b8d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sfa13e100cd7783", "s35620960af7472", "sfb2fda1eb08038", "sf4cfdfab20af6a", "sceeff5dc385420", "sf70484a056af7f", "sfb751107480370", "s0d1c545a72fca0", "s844f17699103d6" ] }, { "age": 100, "six": 47.51166666666666, "five": 49.878, "tasks": { "blimp_filtered": 67.92, "supplement_filtered": 57.98, "ewok_filtered": 51.55, "entity_tracking": 19.38, "comps": 52.56, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 50.0 }, "checkpoint": "81e6ac7bf5074ea166a5450b53120349bed874f46143dee116efaa9f6f9fa21e", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sfa13e100cd7783", "s35620960af7472", "s1f3e387ce03fd7", "s14a990f9be08e1", "sceeff5dc385420", "sf70484a056af7f", "sfb751107480370", "s0d1c545a72fca0", "s844f17699103d6" ] } ], "peak": { "five": { "score": 50.552, "ages": [ 80 ], "rank": 105 }, "six": { "score": 48.32083333333333, "ages": [ 80 ], "rank": 95 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 74, "id": "run/IPRLM384_NORM820_SPREAD640_SEED20260907_40M_20261021", "original_name": "IPRLM384_NORM820_SPREAD640_SEED20260907_40M_20261021", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "Norm820＋FFN分散入力LR約5.26％増", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All21same-seed40Mconditions;all7tasks,5/6aggregates,pairedoutcomes;fullinventory;deltaagainstNorm820parent" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 137, "sources": [ "sad154f1f5e8d5f", "sfd2cb439b52725" ], "points": [ { "age": 40.0, "six": 48.17, "five": 50.474, "tasks": { "blimp_filtered": 68.71, "supplement_filtered": 59.82, "ewok_filtered": 50.0, "entity_tracking": 21.04, "comps": 52.8, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "e2da8d03b36ad21a7b0450eb5233caea9cec759294f52f49525be2133c12ba9d", "precision": "fp32", "verification": "primary", "sources": [ "s92e3fda11d169b", "sbbbd0dad5c46a3" ] }, { "age": 70.0, "six": 48.30583333333333, "five": 50.446, "tasks": { "blimp_filtered": 70.42, "supplement_filtered": 58.87, "ewok_filtered": 51.09, "entity_tracking": 19.52, "comps": 52.33, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 49.0 }, "checkpoint": "f32b3c1ecc7c527ade72db4d0dee1855316cdefbba4a95396e097f8fb16eb5a0", "precision": "fp32", "verification": "primary", "sources": [ "s9f4be038efddfe", "sbbbd0dad5c46a3" ] }, { "age": 80.0, "six": 47.678333333333335, "five": 49.972, "tasks": { "blimp_filtered": 68.19, "supplement_filtered": 59.49, "ewok_filtered": 50.51, "entity_tracking": 19.47, "comps": 52.2, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 53.0 }, "checkpoint": "9ef9caec83463932e9ff344dcf6e1a65ac12c1e997a1096cdd76229d633eeb91", "precision": "fp32", "verification": "primary", "sources": [ "s8e665f8009c2e6", "sbbbd0dad5c46a3" ] } ], "peak": { "five": { "score": 50.474, "ages": [ 40.0 ], "rank": 111 }, "six": { "score": 48.30583333333333, "ages": [ 70.0 ], "rank": 96 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 75, "id": "baby/W_READER64_WIDTH_AWARE", "original_name": "Reader64・幅対応更新", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 64, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 64, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 34, "sources": [ "sea69f62068bf50", "sb90d607114f8b8", "s0b0e008ac94739", "s82a5b28c0a46ea" ], "points": [ { "age": 80, "six": 48.270833333333336, "five": 50.592, "tasks": { "blimp_filtered": 68.12, "supplement_filtered": 60.85, "ewok_filtered": 51.3, "entity_tracking": 20.38, "comps": 52.31, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 51.0 }, "checkpoint": "fde14a2f77f627dad222f9e0a705908fa83b3bbb4e5677b6a27abd5f3d71079a", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s19907c9ef610a6", "s90d0b5bb7b045c", "sbb4d74410c88fd", "s0fdb2e747bd361", "s3513894e36f0db" ] }, { "age": 100, "six": 48.10166666666667, "five": 50.398, "tasks": { "blimp_filtered": 68.47, "supplement_filtered": 60.21, "ewok_filtered": 51.85, "entity_tracking": 19.13, "comps": 52.33, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 48.0 }, "checkpoint": "ab4a16141d28d32a9587d2eb99d76da7b557403e01873c10fdf0c46a6917a69f", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s19907c9ef610a6", "s90d0b5bb7b045c", "sb8971a0c63efd4", "s0fdb2e747bd361", "s3513894e36f0db" ] } ], "peak": { "five": { "score": 50.592, "ages": [ 80 ], "rank": 98 }, "six": { "score": 48.270833333333336, "ages": [ 80 ], "rank": 97 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 76, "id": "baby/B_READER_DEPTH9_WIDTH272_NO_DISTANCE_KNOBS80_INIT_100M", "original_name": "旧272・距離なし・80Mノブ初期値", "title": "272幅・FFN960・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 272, "ffn": 960, "depth": 9, "parts": { "width": 272, "ffn": 960, "depth": 9, "attention": "8×34", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5294117647058822 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 272, "layers": 9, "hidden": 960, "attention_heads": 8, "head_dim": 34, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "history_norm": "separate", "seed": 20260907, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "knob_initialization": "learned256_080M", "knob_seed_sha256": "fba134e1b4647e9c6dee71f43481ef721b1c9afac40862eb3cd9427fc5992344", "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score60/70/80/90/100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 26, "sources": [ "s4888135e9a6e94", "sc6442f82082fca", "s8fb648af057a92", "s93129c03b6c606", "s0f3af6810601e3", "s41d6886102e6a6", "s946cea63ad4dce" ], "points": [ { "age": 60, "six": 46.505, "five": 49.37, "tasks": { "blimp_filtered": 67.77, "supplement_filtered": 57.24, "ewok_filtered": 50.59, "entity_tracking": 19.46, "comps": 51.79, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 43.0 }, "checkpoint": "76178e74da3c36dd5516b0a2ac432579f8a14e6bdf6b47089afa93eb4bbcbb56", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4bcdfa2067e575", "s1029b90d20cb13", "sab1a27c64c21f4", "s88dd348b83d8a3", "sf89d6f793215c4", "s4d9db48c92cd90", "sc13a4040b91508", "seddd64eadd6b89", "s8a98b42ab4acb1" ] }, { "age": 70, "six": 47.50666666666667, "five": 50.166, "tasks": { "blimp_filtered": 69.06, "supplement_filtered": 59.27, "ewok_filtered": 50.61, "entity_tracking": 19.84, "comps": 52.05, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 49.0 }, "checkpoint": "5d8553b62fcee46c32f298f0a8850c3f2150c99cebc1ef83e156021a718d3ff9", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4bcdfa2067e575", "s1029b90d20cb13", "s7be369001604a0", "sd9a4fa1ee2a764", "sf89d6f793215c4", "s4d9db48c92cd90", "sc13a4040b91508", "seddd64eadd6b89", "s8a98b42ab4acb1" ] }, { "age": 80, "six": 48.251666666666665, "five": 50.472, "tasks": { "blimp_filtered": 67.7, "supplement_filtered": 60.63, "ewok_filtered": 51.46, "entity_tracking": 20.52, "comps": 52.05, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "b3d85477d5400fa4564bad2f6bc3e59af76f5f680921806b644b4d8732e4798d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4bcdfa2067e575", "s1029b90d20cb13", "s0da58216408c01", "s222a1aadf218eb", "sf89d6f793215c4", "s4d9db48c92cd90", "sc13a4040b91508", "seddd64eadd6b89", "s8a98b42ab4acb1" ] }, { "age": 90, "six": 47.36333333333334, "five": 49.506, "tasks": { "blimp_filtered": 67.45, "supplement_filtered": 57.35, "ewok_filtered": 51.16, "entity_tracking": 19.62, "comps": 51.95, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "cb9772106c82f23726f2a126596f274f9ce5ecc7f28512954c08ea9834f6bc1a", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4bcdfa2067e575", "s1029b90d20cb13", "s9cb31e4e37f51f", "se5eb947cfd66b8", "sf89d6f793215c4", "s4d9db48c92cd90", "sc13a4040b91508", "seddd64eadd6b89", "s8a98b42ab4acb1" ] }, { "age": 100, "six": 47.964166666666664, "five": 50.33, "tasks": { "blimp_filtered": 68.21, "supplement_filtered": 58.96, "ewok_filtered": 52.53, "entity_tracking": 20.07, "comps": 51.88, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 48.0 }, "checkpoint": "4d2d4444551e15c1cdf403a9e72baddd6a9dc0692d87530d764d51fa22e506d9", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4bcdfa2067e575", "s1029b90d20cb13", "s37f99fbf555275", "s299320b6b4c2f0", "sf89d6f793215c4", "s4d9db48c92cd90", "sc13a4040b91508", "seddd64eadd6b89", "s8a98b42ab4acb1" ] } ], "peak": { "five": { "score": 50.472, "ages": [ 80 ], "rank": 112 }, "six": { "score": 48.251666666666665, "ages": [ 80 ], "rank": 98 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 77, "id": "baby/vistarc_q3_multi_input_mos2_20261004", "original_name": "Q3・MoS2・3層参照・受信ゲート前", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos2_last3", "attention_received_gate": false, "ffn_input_output_gates": false, "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true }, "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "metric": "maximum observed local six-axis", "train_words": 100000000, "save_words": [ 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M training completes; serial", "parent_peak": 48.57583333333333, "historical_record": 49.03583333333333, "record_gap_from_parent": 0.46000000000000085, "no_endpoint_penalty": true, "diagnostics": [ "same-checkpoint five-axis and seven component scores", "training loss", "candidate and reader usage", "mixture usage", "runtime" ], "comparison_scope": "observed record comparison; Q3 currently scored only80/100 so candidate has more scored checkpoints; no equal-search-opportunity causal claim" } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260, "fused_mixture": "eager-equivalent Triton FP32 two-term logsumexp;native broadcast reduction backward" }, "condition_conflicts": [], "profile": 62, "sources": [ "s49564135191237", "sa48a2d0aab330b", "se8f094f2f75d4a", "s6d3970f4a73c3e", "sc33e848e41dd6e", "sd1dd9c179427a7" ], "points": [ { "age": 60, "six": 47.29, "five": 49.412, "tasks": { "blimp_filtered": 67.73, "supplement_filtered": 55.91, "ewok_filtered": 52.27, "entity_tracking": 19.42, "comps": 51.73, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 52.0 }, "checkpoint": "8745ba52c88e4dea04487727d9ce090b603b2a380580e6cb4804256a2b62abd1", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "se819559d799dd9", "s31c50f0b6fbb90", "s6e6e6e4a32600f", "sda648e65ecc0bc", "sc572d80db4e78a", "s524310f0481dc9", "s78878810589388", "s8a473706e42cf0" ] }, { "age": 70, "six": 47.00083333333333, "five": 49.662, "tasks": { "blimp_filtered": 69.77, "supplement_filtered": 56.34, "ewok_filtered": 50.99, "entity_tracking": 19.67, "comps": 51.54, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "824a9b120fde7103695a1c7c911a1472582335f7578d79c4c18715af0bd0f639", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc9a85cdb0c25dd", "s31c50f0b6fbb90", "s85820f8947a52e", "sda648e65ecc0bc", "sc572d80db4e78a", "s524310f0481dc9", "s78878810589388", "s8a473706e42cf0" ] }, { "age": 80, "six": 48.23583333333333, "five": 50.144, "tasks": { "blimp_filtered": 68.64, "supplement_filtered": 59.04, "ewok_filtered": 51.19, "entity_tracking": 19.84, "comps": 52.01, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 57.0 }, "checkpoint": "2ade98e74b7243be0bc25947ea9a7c461e1a7beddb1ca42a70e10a29300a3de9", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1ac6847a970480", "s31c50f0b6fbb90", "s71d9f9c28950ce", "sda648e65ecc0bc", "sc572d80db4e78a", "s524310f0481dc9", "s78878810589388", "s8a473706e42cf0" ] }, { "age": 90, "six": 46.74333333333333, "five": 49.95, "tasks": { "blimp_filtered": 69.33, "supplement_filtered": 57.67, "ewok_filtered": 51.3, "entity_tracking": 19.27, "comps": 52.18, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 42.0 }, "checkpoint": "142b25148103fdeca63bdb618fc9db4955add6f704a3581197388dbab688de71", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s935dedeadeb4a7", "s31c50f0b6fbb90", "se2c29e0b7ca51c", "sda648e65ecc0bc", "sc572d80db4e78a", "s524310f0481dc9", "s78878810589388", "s8a473706e42cf0" ] }, { "age": 100, "six": 47.4075, "five": 50.056, "tasks": { "blimp_filtered": 69.57, "supplement_filtered": 56.77, "ewok_filtered": 50.86, "entity_tracking": 20.82, "comps": 52.26, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "5cf714f0dfedfed024d104862aa0beff042e6dc4cdb1d1737ca0fb382fd02d0c", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "se6d4b90b79f303", "s31c50f0b6fbb90", "sa37d1bfa059a22", "sda648e65ecc0bc", "sc572d80db4e78a", "s524310f0481dc9", "s78878810589388", "s8a473706e42cf0" ] } ], "peak": { "five": { "score": 50.144, "ages": [ 80 ], "rank": 139 }, "six": { "score": 48.23583333333333, "ages": [ 80 ], "rank": 99 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 78, "id": "run/IPRLM384_FFN1408_MAJOR95_OTHER105_SEED20260907_40M_20261020", "original_name": "IPRLM384_FFN1408_MAJOR95_OTHER105_SEED20260907_40M_20261020", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要−5%＋その他＋5%", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005066666666666667, 0.0004433333333333333, 0.00037999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Newjoint+same-age7existingLRconditions;fullinventory;no final80M ranking from40M" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 123, "sources": [ "sb6894b49d1ae64", "scfca9995748dca" ], "points": [ { "age": 40.0, "six": 48.235, "five": 50.252, "tasks": { "blimp_filtered": 68.49, "supplement_filtered": 59.27, "ewok_filtered": 52.02, "entity_tracking": 19.29, "comps": 52.19, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 53.0 }, "checkpoint": "82c24a1288bf999c4304b263e3c3d6ae34cd98b124a388b5318b64c0ed5e9498", "precision": "fp32", "verification": "primary", "sources": [ "s5e09def1e47d23", "s75bcfdaab9ee4b" ] } ], "peak": { "five": { "score": 50.252, "ages": [ 40.0 ], "rank": 131 }, "six": { "score": 48.235, "ages": [ 40.0 ], "rank": 100 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 79, "id": "baby/vistarc_q3_mos3_10l_received_value_gate_20261006", "original_name": "256・10層・MoS3・3層混合", "title": "256幅・FFN896・10層", "topology": "MoS3", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 10, "parts": { "width": 256, "ffn": 896, "depth": 10, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 10, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last3", "attention_received_gate": true, "output_components": 3, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3_10L run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "metric": "maximum observed local six-axis", "train_words": 100000000, "save_words": [ 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "benchmark_words": [ 80000000, 100000000 ], "timing": "after100M training completes; serial", "historical_record": 49.03583333333333, "no_endpoint_penalty": true, "diagnostics": [ "same-checkpoint five-axis and seven component scores", "training loss", "candidate and reader usage", "mixture usage", "runtime" ], "comparison_scope": "10L MoS3+received gate versus original9L MoS3 at70/80/90/100M; target maximum six axes >50", "primary": "Maximum observed local six-axis across70/80/90/100M", "target_six_axis": 50.0, "parent_peak": 49.2675, "terminal_drop_not_rejection": true } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 42, "sources": [ "sc6565b35c025c4", "s05a5e41fbc0891", "sb3380ff1fdd2ae", "sf73e23afd03a4a", "sc2efbdfe26a628", "s52bbc33532e400", "s8383a462b13c82" ], "points": [ { "age": 70, "six": 48.0925, "five": 50.678, "tasks": { "blimp_filtered": 69.5, "supplement_filtered": 58.78, "ewok_filtered": 51.51, "entity_tracking": 21.05, "comps": 52.55, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 48.0 }, "checkpoint": "841df821baf0ddcf6077a4bfa6d2f44b079fdd1a5d0e9b57d70a762f43ef4e23", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc416a854fd50e6", "sf3342a4edf72b4", "s999ec34888212c", "s1c1923961bd2e4", "s6bed34eef3b1f6", "scc3f422b47ea87", "s244db0829a2dc0" ] }, { "age": 80, "six": 47.79083333333333, "five": 50.21, "tasks": { "blimp_filtered": 67.72, "supplement_filtered": 60.22, "ewok_filtered": 51.23, "entity_tracking": 19.68, "comps": 52.2, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 51.0 }, "checkpoint": "f0dfa90fbd5322ceb150366246da14dbbc063ab812ef26337d1629b686bd884e", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s9ab446df5a08a6", "sf3342a4edf72b4", "s212dd6c116b07c", "s1c1923961bd2e4", "s6bed34eef3b1f6", "scc3f422b47ea87", "s244db0829a2dc0" ] }, { "age": 90, "six": 47.105, "five": 49.99, "tasks": { "blimp_filtered": 67.33, "supplement_filtered": 58.55, "ewok_filtered": 51.07, "entity_tracking": 20.37, "comps": 52.63, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 44.0 }, "checkpoint": "18cd46597bf7989f7caae9adf9277c0871213c3d31bc9177741ca10350e2d0b2", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s73dad55ab92a77", "sf3342a4edf72b4", "s91842d2c614234", "s1c1923961bd2e4", "s6bed34eef3b1f6", "scc3f422b47ea87", "s244db0829a2dc0" ] }, { "age": 100, "six": 48.23416666666667, "five": 50.254, "tasks": { "blimp_filtered": 69.0, "supplement_filtered": 58.36, "ewok_filtered": 50.99, "entity_tracking": 20.45, "comps": 52.47, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 52.0 }, "checkpoint": "1dbbd49ab96df09cc232f0afb5c5cf270db1b7ca2f3cf80d9d084c28f44c1eb0", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sa9e1aa219de1bc", "sf3342a4edf72b4", "s095085b3ec9f9e", "s1c1923961bd2e4", "s6bed34eef3b1f6", "scc3f422b47ea87", "s244db0829a2dc0" ] } ], "peak": { "five": { "score": 50.678, "ages": [ 70 ], "rank": 85 }, "six": { "score": 48.23416666666667, "ages": [ 100 ], "rank": 101 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 80, "id": "run/IPRLM336_FFN1176_ADDITIVE_SEED20260907_20261016", "original_name": "IPRLM336_FFN1176_ADDITIVE_SEED20260907_20261016", "title": "336幅・FFN1176・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1176, "depth": 9, "parts": { "width": 336, "ffn": 1176, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 336, "layers": 9, "hidden": 1176, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed same40M parent320 and additive320; final record49.8675 remains separate" }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 90, "sources": [ "s9176b74478a387", "sd143f64d4c2412", "s57d91c383b291f" ], "points": [ { "age": 40.0, "six": 48.22833333333333, "five": 50.832, "tasks": { "blimp_filtered": 69.47, "supplement_filtered": 59.7, "ewok_filtered": 52.12, "entity_tracking": 20.25, "comps": 52.62, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 51.0 }, "checkpoint": "00ac77b9986bd2cef0a053fd4e3f6e552d7d07e1a86c75004a2f2c22dd015dcf", "precision": "fp32", "verification": "primary", "sources": [ "s2e0ed86e56ff71", "s5b62322841f369" ] } ], "peak": { "five": { "score": 50.832, "ages": [ 40.0 ], "rank": 70 }, "six": { "score": 48.22833333333333, "ages": [ 40.0 ], "rank": 102 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 81, "id": "run/IPRLM320_SEED731946281_20261013", "original_name": "IPRLM320_SEED731946281_20261013", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000 at70/80/90/100M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.715, "comparison_scope": "Same parent320 architecture/LR/data order; one alternate seed; common70/80/90/100M peak and individual tasks. Two runs do not estimate seed variance precisely." }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 88, "sources": [ "s9d4d8b6f63fb84", "s7fa625d77b8bcb", "sc470a5c86c5c11" ], "points": [ { "age": 40.0, "six": 47.29333333333334, "five": 50.11, "tasks": { "blimp_filtered": 68.91, "supplement_filtered": 57.47, "ewok_filtered": 52.0, "entity_tracking": 19.63, "comps": 52.54, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 47.0 }, "checkpoint": "7cde3efd56634c516f654dfbca2203bd553e127d769068df1964faf39774c4a7", "precision": "fp32", "verification": "primary", "sources": [ "sd30ea29eddaf4f", "sff86268a78b7fe" ] }, { "age": 70.0, "six": 48.214166666666664, "five": 50.718, "tasks": { "blimp_filtered": 69.57, "supplement_filtered": 58.43, "ewok_filtered": 52.05, "entity_tracking": 20.99, "comps": 52.55, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 51.0 }, "checkpoint": "4cf1b70e26872e34fd39ed5a0ce4993725646aa4e4a2b0bee788bb9c1c8a1755", "precision": "fp32", "verification": "primary", "sources": [ "s0b1dd5fe35304b", "s39689049b76b57", "sbebf3544f7b5cd", "s17e296c8ef08fa" ] }, { "age": 80.0, "six": 48.090833333333336, "five": 50.276, "tasks": { "blimp_filtered": 67.97, "supplement_filtered": 59.43, "ewok_filtered": 51.69, "entity_tracking": 19.9, "comps": 52.39, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "checkpoint": "c4771d7651f3d4e20a24816e39e8794e19532b4d6858a144d33a0b3ce06bf2d1", "precision": "fp32", "verification": "primary", "sources": [ "s60b76af63963e6", "s815639cce73b99", "sbebf3544f7b5cd", "s17e296c8ef08fa" ] }, { "age": 90.0, "six": 47.64666666666667, "five": 49.452, "tasks": { "blimp_filtered": 67.98, "supplement_filtered": 57.0, "ewok_filtered": 51.16, "entity_tracking": 19.42, "comps": 51.7, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 52.0 }, "checkpoint": "b11a5e4f2c76152c099dde29d6c95656c6e1be4be3cb136930aedcc111e8c0dc", "precision": "fp32", "verification": "primary", "sources": [ "s394a2604492494", "se87e28a21715e3", "sbebf3544f7b5cd", "s17e296c8ef08fa" ] }, { "age": 100.0, "six": 47.82666666666667, "five": 50.362, "tasks": { "blimp_filtered": 68.26, "supplement_filtered": 58.57, "ewok_filtered": 52.35, "entity_tracking": 20.01, "comps": 52.62, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 47.0 }, "checkpoint": "eb1e19fd285ee334146c777a5435b0cf9a4d5d15ba4a759b091d884086cd4b1f", "precision": "fp32", "verification": "primary", "sources": [ "s3a9222a3254050", "s15b2596b66bab0", "sbebf3544f7b5cd", "s17e296c8ef08fa" ] } ], "peak": { "five": { "score": 50.718, "ages": [ 70.0 ], "rank": 82 }, "six": { "score": 48.214166666666664, "ages": [ 70.0 ], "rank": 103 } }, "clear_points": [ { "number": 81, "age": 70, "seed": 731946281, "checkpoint_sha256": "4cf1b70e26872e34fd39ed5a0ce4993725646aa4e4a2b0bee788bb9c1c8a1755", "old_six": 48.214166666666664, "old_five": 50.718, "old_tasks": { "blimp_filtered": 69.57, "supplement_filtered": 58.43, "ewok_filtered": 52.05, "entity_tracking": 20.99, "comps": 52.55, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 51.0 }, "correct": 12389, "total": 17274, "accuracy": 71.72050480490911, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6989, "total": 8392, "ties": 0 }, "comps": { "correct": 3998, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 1, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 12, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 0, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 988, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 4, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 4, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 5, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 73, "total": 114, "ties": 0 }, "boolq": { "correct": 9, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 227, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 62.55751232599869, "checkpoint_rank": 3, "rank": 3 }, { "number": 81, "age": 80, "seed": 731946281, "checkpoint_sha256": "c4771d7651f3d4e20a24816e39e8794e19532b4d6858a144d33a0b3ce06bf2d1", "old_six": 48.090833333333336, "old_five": 50.276, "old_tasks": { "blimp_filtered": 67.97, "supplement_filtered": 59.43, "ewok_filtered": 51.69, "entity_tracking": 19.9, "comps": 52.39, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "correct": 12191, "total": 17274, "accuracy": 70.57427347458608, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6838, "total": 8392, "ties": 0 }, "comps": { "correct": 3803, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 3, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 15, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1003, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 1, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 1, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 0, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 73, "total": 114, "ties": 0 }, "boolq": { "correct": 10, "total": 53, "ties": 0 }, "mnli": { "correct": 28, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 358, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 63.04400519900082, "checkpoint_rank": 15 } ], "clear_peak": { "number": 81, "age": 70, "seed": 731946281, "checkpoint_sha256": "4cf1b70e26872e34fd39ed5a0ce4993725646aa4e4a2b0bee788bb9c1c8a1755", "old_six": 48.214166666666664, "old_five": 50.718, "old_tasks": { "blimp_filtered": 69.57, "supplement_filtered": 58.43, "ewok_filtered": 52.05, "entity_tracking": 20.99, "comps": 52.55, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 51.0 }, "correct": 12389, "total": 17274, "accuracy": 71.72050480490911, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6989, "total": 8392, "ties": 0 }, "comps": { "correct": 3998, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 1, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 12, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 0, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 988, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 4, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 4, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 5, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 73, "total": 114, "ties": 0 }, "boolq": { "correct": 9, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 227, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 62.55751232599869, "checkpoint_rank": 3, "rank": 3 }, "present": true }, { "number": 82, "id": "baby/vistarc_q3_mos3_last2_width240_ffn832_attn256_20261007", "original_name": "240・MoS3・2層混合・Attention256", "title": "240幅・FFN832・9層", "topology": "MoS3・共通2層混合", "recipe": "主要0.0008／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 240, "ffn": 832, "depth": 9, "parts": { "width": 240, "ffn": 832, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 8, 9 ] ], "ffn_ratio": 3.466666666666667 }, "main_lr": [ 0.0008, 0.0007, 0.0006 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 240, "layers": 9, "hidden": 832, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "fan_in_init_only", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last2", "attention_received_gate": true, "output_components": 3, "output_state_layers": 2, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3_LAST2 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 65536, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.315, "comparison_scope": "last2_256 parent plus interrupted last2_272 and last3/last4 references" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 49, "sources": [ "sd3c864adbba8f4", "sd4fe6414422456", "s8fe8a8ef8e8d69", "sbc521993835e27", "s84c333dae8d544", "s47cc89907ccb20" ], "points": [ { "age": 70, "six": 47.97083333333333, "five": 50.144, "tasks": { "blimp_filtered": 69.83, "supplement_filtered": 56.55, "ewok_filtered": 51.14, "entity_tracking": 20.64, "comps": 52.56, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 48.0 }, "checkpoint": "8b7c82946e9ca7e174c4ec383aa665e6bdd8fa51ff2f8efb02c635aba15099cd", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s569ba59a8e57db", "sc7845ba4f5a20c", "s5ad34e0b12a2bc", "s8cd4515e93e0c3", "s0425e183883480", "s8335b3e3963e1e" ] }, { "age": 80, "six": 48.20916666666667, "five": 50.212, "tasks": { "blimp_filtered": 68.48, "supplement_filtered": 58.96, "ewok_filtered": 51.52, "entity_tracking": 19.49, "comps": 52.61, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 56.0 }, "checkpoint": "4dadf1164f7362083311b9fca6c4d8d07cb19586b351673eeb57eb8a16515dab", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "scff6d435f80eb1", "s2909b90e076558", "s5ad34e0b12a2bc", "s8cd4515e93e0c3", "s0425e183883480", "s8335b3e3963e1e" ] }, { "age": 90, "six": 47.34916666666667, "five": 49.492, "tasks": { "blimp_filtered": 68.24, "supplement_filtered": 56.81, "ewok_filtered": 51.05, "entity_tracking": 18.71, "comps": 52.65, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 49.0 }, "checkpoint": "24d877f51bcf5df0b5fe2c019375993e2eb98e561a76e283fd2e6436ca118a3f", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s2ae0b228598690", "s71e6981fc66814", "s5ad34e0b12a2bc", "s8cd4515e93e0c3", "s0425e183883480", "s8335b3e3963e1e" ] }, { "age": 100, "six": 48.02, "five": 50.394, "tasks": { "blimp_filtered": 68.97, "supplement_filtered": 58.59, "ewok_filtered": 50.49, "entity_tracking": 21.59, "comps": 52.33, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 49.0 }, "checkpoint": "34d3fe7fb8c9c6ecbc946fb81ffae1bb307972efb270efca59e2b2c220d7843b", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s770b56c65f6b39", "s6c63ac2c50878f", "s5ad34e0b12a2bc", "s8cd4515e93e0c3", "s0425e183883480", "s8335b3e3963e1e" ] } ], "peak": { "five": { "score": 50.394, "ages": [ 100 ], "rank": 120 }, "six": { "score": 48.20916666666667, "ages": [ 80 ], "rank": 104 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 83, "id": "baby/vistarc_q3_mos3_self_review_20261007", "original_name": "256・MoS3・3層混合＋自己見直し", "title": "256幅・FFN896・9層", "topology": "MoS3・共通3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last3_review", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "review_width": 128, "judge_width": 64, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "metric": "highest observed local six-axis", "train_words": 100000000, "no_endpoint_penalty": true, "parent_peak": 49.2675, "series_record": 49.315, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "diagnostics": [ "five axes and each task", "per-question predictions", "draft/proposal/selection gains", "rescue and damage", "judge calibration" ] }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 58, "sources": [ "s19d26a2edfabc3", "s638b1c6a405f7a", "sc64d104fff83ed", "saf59dff67c5c72", "sb787af4542f1ac", "safdc2487c2aca0" ], "points": [ { "age": 70, "six": 48.16833333333334, "five": 50.366, "tasks": { "blimp_filtered": 69.39, "supplement_filtered": 58.93, "ewok_filtered": 52.02, "entity_tracking": 19.69, "comps": 51.8, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 53.0 }, "checkpoint": "d6101006baee07a2efefba42f812ef39a74e00e992f6f12dc1cc0fb40a69e8b1", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s84f563074a2312", "s67e35caa5e87ff", "s4cc3f1eb832d4e", "se2806658cba876", "s41773e82cad24d", "s5052d878720ad2" ] }, { "age": 80, "six": 48.18, "five": 50.486, "tasks": { "blimp_filtered": 68.69, "supplement_filtered": 59.24, "ewok_filtered": 52.68, "entity_tracking": 19.67, "comps": 52.15, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "a2de41e0fcfb592d1d3b8815aa31d79dc4bc37bb677f7388aac39d3b73b4e14a", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sb0f1e205bd901f", "sdf59ad6b25b4d0", "s4cc3f1eb832d4e", "se2806658cba876", "s41773e82cad24d", "s5052d878720ad2" ] }, { "age": 90, "six": 47.6475, "five": 50.144, "tasks": { "blimp_filtered": 69.67, "supplement_filtered": 57.1, "ewok_filtered": 52.5, "entity_tracking": 19.37, "comps": 52.08, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 48.0 }, "checkpoint": "a318a11af8fdabf4b0db35db5ef62d6d39962c5f0e4e6d975162b7906e97f6cf", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "scce8b410c4fafe", "sa8ba62bcf2c228", "s4cc3f1eb832d4e", "se2806658cba876", "s41773e82cad24d", "s5052d878720ad2" ] }, { "age": 100, "six": 47.111666666666665, "five": 50.098, "tasks": { "blimp_filtered": 69.48, "supplement_filtered": 57.15, "ewok_filtered": 51.85, "entity_tracking": 19.63, "comps": 52.38, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 43.0 }, "checkpoint": "bec0e93fbf7a6f8e024b4796b110a5d94450842c2c61f4099d2fa02e439845fe", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sbcd2b3aaef24ff", "s1e235a22d34f66", "s4cc3f1eb832d4e", "se2806658cba876", "s41773e82cad24d", "s5052d878720ad2" ] } ], "peak": { "five": { "score": 50.486, "ages": [ 80 ], "rank": 109 }, "six": { "score": 48.18, "ages": [ 80 ], "rank": 105 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 84, "id": "baby/B_READER_DEPTH9_JOINT_DELTA_100M", "original_name": "旧256・9層・差分共同正規化", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "未確認", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "history_norm": "joint_pair", "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 17, "sources": [ "se88c16876f1e2a", "s95629bc1cd810b", "s897ad635cfa559", "s53ab8e873a57da", "se8b8619481b9d8", "sa7b25a98d8d4ef", "sb2d813f2655b1d" ], "points": [ { "age": 80, "six": 48.16916666666667, "five": 50.464, "tasks": { "blimp_filtered": 68.57, "supplement_filtered": 58.95, "ewok_filtered": 51.92, "entity_tracking": 20.38, "comps": 52.5, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 53.0 }, "checkpoint": "3da15d28e63b634a026a01f2caed38ff38b7c2035fc4053be66bf3d5872ea0c9", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc9940ad6a6e1dd", "s737e3161c014dd", "s6d9a4adece3061", "s3e154c3b340862", "s0764806b26e826", "s50d16ddb7408a8", "s88f34cf1d28667", "s1a0ac36fc5b12d", "s217f26afda0151" ] }, { "age": 100, "six": 47.19916666666666, "five": 49.906, "tasks": { "blimp_filtered": 68.32, "supplement_filtered": 57.55, "ewok_filtered": 52.18, "entity_tracking": 19.74, "comps": 51.74, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 45.0 }, "checkpoint": "35ff93f5ec31cb87c0b9633ff4f004c848ce44323878655daf22881d6d1775f5", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc9940ad6a6e1dd", "s737e3161c014dd", "s7db651d837b56c", "sd68e6275d0bac5", "s0764806b26e826", "s50d16ddb7408a8", "s88f34cf1d28667", "s1a0ac36fc5b12d", "s217f26afda0151" ] } ], "peak": { "five": { "score": 50.464, "ages": [ 80 ], "rank": 115 }, "six": { "score": 48.16916666666667, "ages": [ 80 ], "rank": 106 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 85, "id": "baby/Q0_READER32_REFERENCE", "original_name": "Q0・Reader32 Reference", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 30, "sources": [ "s8ba2bb0b00bc9e", "s6fa269501b0968", "s56d4df24d4577e", "se5879a614752f4", "s1029cbc5432b9b" ], "points": [ { "age": 80, "six": 47.37166666666667, "five": 50.104, "tasks": { "blimp_filtered": 67.97, "supplement_filtered": 59.36, "ewok_filtered": 52.79, "entity_tracking": 18.08, "comps": 52.32, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 48.0 }, "checkpoint": "2414b49a490b4c5b8870979873746a8005f4684ecb6fa555f6b704aaf1f8d27f", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sedc9e169d5ed51", "s260c4a546380db", "s3b33d09b5bcb22", "s4b50213b09a241", "s8c5cdf92f9669d", "s3e080b973c470f", "s3eacae43204156" ] }, { "age": 100, "six": 48.13666666666666, "five": 51.014, "tasks": { "blimp_filtered": 68.84, "supplement_filtered": 59.88, "ewok_filtered": 53.43, "entity_tracking": 20.47, "comps": 52.45, "global_piqa_parallel": 16.5, "global_piqa_nonparallel": 51.0 }, "checkpoint": "4814b29d858311d0819f2c64c3078854b928efd86eaab2b9af23bbc55087ff3f", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sedc9e169d5ed51", "s260c4a546380db", "s6b2b0444b7db32", "s80f49b7a2d84ed", "s8c5cdf92f9669d", "s3e080b973c470f", "s3eacae43204156" ] } ], "peak": { "five": { "score": 51.014, "ages": [ 100 ], "rank": 44 }, "six": { "score": 48.13666666666666, "ages": [ 100 ], "rank": 107 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 86, "id": "run/IPRLM384_A_EMBEDDING_MINUS5_SEED20260907_40M_20261021", "original_name": "IPRLM384_A_EMBEDDING_MINUS5_SEED20260907_40M_20261021", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "A＋入力embeddingだけ−5％", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005066666666666667, 0.0004433333333333333, 0.00037999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All18same-seed40M LRconditions;reuse15controls;allgenres/paired/fullinventory" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 96, "sources": [ "s8598d0f0816999", "sb3a08e60424fcd" ], "points": [ { "age": 40.0, "six": 48.09916666666667, "five": 50.292, "tasks": { "blimp_filtered": 68.72, "supplement_filtered": 59.23, "ewok_filtered": 51.41, "entity_tracking": 20.3, "comps": 51.8, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 50.0 }, "checkpoint": "ee5b4e226dbf8e5d82b1ee965a6002ac09cac595d2f79e170f77025d3f4ee6d2", "precision": "fp32", "verification": "primary", "sources": [ "s48e45a718d58f0", "s1f55674c641d91" ] } ], "peak": { "five": { "score": 50.292, "ages": [ 40.0 ], "rank": 125 }, "six": { "score": 48.09916666666667, "ages": [ 40.0 ], "rank": 108 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 87, "id": "baby/vistarc_q3_mos3_all9_dynamic_20261006", "original_name": "256・MoS3・全9層・入力別割合", "title": "256幅・FFN896・9層", "topology": "MoS3・共通9層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "1・2・3・4・5・6・7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 1, 2, 3, 4, 5, 6, 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_all9", "attention_received_gate": true, "output_components": 3, "output_state_layers": 9, "layer_weight_mode": "dynamic", "layer_router_dim": 32, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3_ALL9 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "metric": "maximum observed local six-axis", "train_words": 100000000, "save_words": [ 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M training completes; serial", "historical_record": 49.315, "no_endpoint_penalty": true, "diagnostics": [ "same-checkpoint five-axis and seven component scores", "training loss", "candidate and reader usage", "mixture usage", "runtime" ], "comparison_scope": "All9static vs all9dynamic, plus existing2/3/4/5 state cohorts" } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 43, "sources": [ "se69cbebb4a96fa", "s785591a9e63710", "s19987f0d251132", "s4370801751baa8", "s6e3326c174a571", "s73910a55e58a33", "s9a62b57de8d0ea" ], "points": [ { "age": 70, "six": 48.090833333333336, "five": 50.376, "tasks": { "blimp_filtered": 70.18, "supplement_filtered": 58.62, "ewok_filtered": 51.76, "entity_tracking": 18.93, "comps": 52.39, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 51.0 }, "checkpoint": "f5be748e0816855c8f5334ea06bea29cc66e41914a7a66b29de3909af45dcde8", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s2e05f535757964", "s1f048170f09d72", "s9280f9d5a6e7bf", "sd3ea3f08121a98", "s55cfbf50df23ec", "s38b8db7db4b89a" ] }, { "age": 80, "six": 47.95666666666666, "five": 50.206, "tasks": { "blimp_filtered": 68.41, "supplement_filtered": 58.57, "ewok_filtered": 52.21, "entity_tracking": 19.48, "comps": 52.36, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 54.0 }, "checkpoint": "7caf47d5c5bae4f06a381738001b8c4b014a6acd01bab1429a1bf24f2f6f9ecf", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sd3f71d63893a6e", "s1f048170f09d72", "s9280f9d5a6e7bf", "sd3ea3f08121a98", "s55cfbf50df23ec", "s38b8db7db4b89a" ] }, { "age": 90, "six": 47.35666666666667, "five": 49.698, "tasks": { "blimp_filtered": 68.28, "supplement_filtered": 59.06, "ewok_filtered": 50.51, "entity_tracking": 18.1, "comps": 52.54, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 48.0 }, "checkpoint": "2dc9204e7aa3fa496277898a744d512dd409cde0c568a081fe9104e5f18d39df", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s493239e2996ee1", "s1f048170f09d72", "s9280f9d5a6e7bf", "sd3ea3f08121a98", "s55cfbf50df23ec", "s38b8db7db4b89a" ] }, { "age": 100, "six": 47.744166666666665, "five": 50.648, "tasks": { "blimp_filtered": 70.16, "supplement_filtered": 60.02, "ewok_filtered": 51.05, "entity_tracking": 19.13, "comps": 52.88, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 48.0 }, "checkpoint": "851e00140f41f70304f0eb88562a1ceedfa7d2b6efcf9d6d65446bd9bf6b2e5d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s449259494fb456", "s1f048170f09d72", "s9280f9d5a6e7bf", "sd3ea3f08121a98", "s55cfbf50df23ec", "s38b8db7db4b89a" ] } ], "peak": { "five": { "score": 50.648, "ages": [ 100 ], "rank": 88 }, "six": { "score": 48.090833333333336, "ages": [ 70 ], "rank": 109 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 88, "id": "run/IPRLM336_FFN1232_ADDITIVE_SEED137_20261016", "original_name": "IPRLM336_FFN1232_ADDITIVE_SEED137_20261016", "title": "336幅・FFN1232・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M parent320/additive320/additive336FFN1176 and siblingFFN variants; observed later peaks separate." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 91, "sources": [ "sa6dc6faabde83c", "scd9b319c0156c0", "sf0d43cc323cf81" ], "points": [ { "age": 40.0, "six": 48.089166666666664, "five": 50.38, "tasks": { "blimp_filtered": 69.3, "supplement_filtered": 58.09, "ewok_filtered": 51.87, "entity_tracking": 20.38, "comps": 52.26, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 49.0 }, "checkpoint": "2a4f25a440aa02a518d3dc8c4092ee5c0056a7d33f89b698ccb5870fb7fd7833", "precision": "fp32", "verification": "primary", "sources": [ "s16606178916571", "s6dc2c7d827fd46" ] } ], "peak": { "five": { "score": 50.38, "ages": [ 40.0 ], "rank": 121 }, "six": { "score": 48.089166666666664, "ages": [ 40.0 ], "rank": 110 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 89, "id": "run/IPRLM320_FFN1280_MATCHED_PARENT_20261012", "original_name": "IPRLM320_FFN1280_MATCHED_PARENT_20261012", "title": "320幅・FFN1280・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1280, "depth": 9, "parts": { "width": 320, "ffn": 1280, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 4.0 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1280, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at70/80/90/100M", "SyntaxGym31 and originalPIQA1838 at70/80/90/100M", "frozen bAbI1947 and ProofWriter1000 at70/80/90/100M" ], "intermediate": "No evaluation during training; all benchmarks after100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Original parent320 and child matched70/80/90/100M, with observed peaks separately" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 76, "sources": [ "s973edbef7b55dc", "s800ec87ccda07c", "s4b095a27d76f69", "s1262c84f27f20f", "s05ca10cf73a9a9", "s21135974d5644b" ], "points": [ { "age": 40.0, "six": 47.75666666666667, "five": 49.792, "tasks": { "blimp_filtered": 68.5, "supplement_filtered": 57.28, "ewok_filtered": 50.69, "entity_tracking": 20.18, "comps": 52.31, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 47.0 }, "checkpoint": "4c40053c888cbfeb0bc3193f8caafe258dc6d817ca6a16138706d8dd00fe80c2", "precision": "fp32", "verification": "primary", "sources": [ "s5f99b29353e7af", "s6af7393dc37521" ] }, { "age": 70.0, "six": 48.02, "five": 50.7, "tasks": { "blimp_filtered": 70.49, "supplement_filtered": 58.17, "ewok_filtered": 52.1, "entity_tracking": 20.42, "comps": 52.32, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 44.0 }, "checkpoint": "c172b1486dcabe9b5d2edce2877840097e2253eb7a7b47b67d1cc66582b1f8b1", "precision": "fp32", "verification": "primary", "sources": [ "scc32ab39c5f5b2", "s64bb3cdf0e2fd9", "s3bbc634e1ab067", "se316c7c3a244bf", "s2cf92bf9510b7f", "sa2a2aa4bd53db5" ] }, { "age": 80.0, "six": 48.08, "five": 50.746, "tasks": { "blimp_filtered": 68.7, "supplement_filtered": 61.14, "ewok_filtered": 51.13, "entity_tracking": 20.4, "comps": 52.36, "global_piqa_parallel": 16.5, "global_piqa_nonparallel": 53.0 }, "checkpoint": "46af10c3b1924d8fb2334486c67f1fb895c693fe41fa5fdeb0d708f3543943a9", "precision": "fp32", "verification": "primary", "sources": [ "s9187ae8133f561", "s182c0776a79aac", "s3bbc634e1ab067", "se316c7c3a244bf", "s2cf92bf9510b7f", "sa2a2aa4bd53db5" ] }, { "age": 90.0, "six": 48.065, "five": 49.962, "tasks": { "blimp_filtered": 69.74, "supplement_filtered": 57.38, "ewok_filtered": 50.51, "entity_tracking": 19.97, "comps": 52.21, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 49.0 }, "checkpoint": "c4e003cc982ba88e80413784e26a80aedc24e8cae66822c924963c3e0d61f0a3", "precision": "fp32", "verification": "primary", "sources": [ "s05670f6db6a6c8", "s35088a92316651", "s3bbc634e1ab067", "se316c7c3a244bf", "s2cf92bf9510b7f", "sa2a2aa4bd53db5" ] }, { "age": 100.0, "six": 47.865833333333335, "five": 50.206, "tasks": { "blimp_filtered": 68.69, "supplement_filtered": 60.02, "ewok_filtered": 50.75, "entity_tracking": 19.39, "comps": 52.18, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 50.0 }, "checkpoint": "6f9811b026dd4efaba73d1af19f3899db383331881427d0b1c2cecc9430aed0f", "precision": "fp32", "verification": "primary", "sources": [ "s849722dae85e51", "s253dc496996bcb", "s3bbc634e1ab067", "se316c7c3a244bf", "s2cf92bf9510b7f", "sa2a2aa4bd53db5" ] } ], "peak": { "five": { "score": 50.746, "ages": [ 80.0 ], "rank": 81 }, "six": { "score": 48.08, "ages": [ 80.0 ], "rank": 111 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 90, "id": "baby/vistarc_q3_mos3_all9_static_20261006", "original_name": "256・MoS3・全9層・共通割合", "title": "256幅・FFN896・9層", "topology": "MoS3・共通9層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "1・2・3・4・5・6・7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 1, 2, 3, 4, 5, 6, 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_all9", "attention_received_gate": true, "output_components": 3, "output_state_layers": 9, "layer_weight_mode": "static", "layer_router_dim": 32, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3_ALL9 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "metric": "maximum observed local six-axis", "train_words": 100000000, "save_words": [ 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M training completes; serial", "historical_record": 49.315, "no_endpoint_penalty": true, "diagnostics": [ "same-checkpoint five-axis and seven component scores", "training loss", "candidate and reader usage", "mixture usage", "runtime" ], "comparison_scope": "All9static vs all9dynamic, plus existing2/3/4/5 state cohorts" } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 43, "sources": [ "s2018f13a00a446", "sb623e52f6306c1", "s6732ca4c5cd922", "s90185365a574d9", "sbc42aef262d411", "s7ddf9620167de1", "s9923bd5bf26c47" ], "points": [ { "age": 70, "six": 48.06166666666667, "five": 50.638, "tasks": { "blimp_filtered": 69.5, "supplement_filtered": 58.63, "ewok_filtered": 51.57, "entity_tracking": 21.35, "comps": 52.14, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 49.0 }, "checkpoint": "f8a7551780551ff8a0e865ca931f54c3f1787089c483d679bb66f234afec511f", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s3cf74e1af09fef", "s1dba92974a9ca1", "sc73b92e4b19242", "s93efe5d6bf7c2b", "sf979eaa9c7d448", "s7a3fffd204cf3b" ] }, { "age": 80, "six": 47.920833333333334, "five": 50.566, "tasks": { "blimp_filtered": 68.56, "supplement_filtered": 59.02, "ewok_filtered": 51.2, "entity_tracking": 21.85, "comps": 52.2, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 49.0 }, "checkpoint": "5cbc7e0a97dca8ec69a45ceefd50800c9423c56e71cb019a973dee621c07c526", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1d45c32e0cc5eb", "s1dba92974a9ca1", "sc73b92e4b19242", "s93efe5d6bf7c2b", "sf979eaa9c7d448", "s7a3fffd204cf3b" ] }, { "age": 90, "six": 47.50833333333333, "five": 49.68, "tasks": { "blimp_filtered": 67.87, "supplement_filtered": 57.37, "ewok_filtered": 50.73, "entity_tracking": 20.08, "comps": 52.35, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "74ea95128d0c95370059b038244c5e3a0b85d00e18379c0c11f8c3fe1a6f4ee8", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s7fe4c05f23e434", "s1dba92974a9ca1", "sc73b92e4b19242", "s93efe5d6bf7c2b", "sf979eaa9c7d448", "s7a3fffd204cf3b" ] }, { "age": 100, "six": 47.93833333333333, "five": 50.59, "tasks": { "blimp_filtered": 69.29, "supplement_filtered": 59.45, "ewok_filtered": 51.88, "entity_tracking": 20.02, "comps": 52.31, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 48.0 }, "checkpoint": "7b2e9994d928bbef05cbcfa57e4dcc0380e0e0bfeb87fb6ac6c102264d776f79", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s17941a7cd9a8a7", "s1dba92974a9ca1", "sc73b92e4b19242", "s93efe5d6bf7c2b", "sf979eaa9c7d448", "s7a3fffd204cf3b" ] } ], "peak": { "five": { "score": 50.638, "ages": [ 70 ], "rank": 89 }, "six": { "score": 48.06166666666667, "ages": [ 70 ], "rank": 112 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 91, "id": "baby/Q4_READER40_CONSERVATIVE", "original_name": "Q4・Reader40 Conservative", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 40, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 40, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 32, "sources": [ "s38debe5019f980", "s23a5eb7e3dd6d7", "s2d4ac121de7661", "sb81c52d76d71a6", "sa6d4766588c1c8" ], "points": [ { "age": 80, "six": 48.05583333333333, "five": 50.44, "tasks": { "blimp_filtered": 68.29, "supplement_filtered": 60.37, "ewok_filtered": 52.01, "entity_tracking": 19.17, "comps": 52.36, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 48.0 }, "checkpoint": "1900051e08a4e0dc0d43cab5c45c8a23a24040ac217282bcf6ebb445b341e887", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sf20cfcad913002", "s0a314e240c8a61", "sd0dacd3290d922", "s67fc7185cbea2f", "s9d09e028688960", "s0cc0331c90aa6e", "sa85b49a1125fe0" ] }, { "age": 100, "six": 47.68333333333333, "five": 50.396, "tasks": { "blimp_filtered": 68.47, "supplement_filtered": 59.25, "ewok_filtered": 52.38, "entity_tracking": 19.14, "comps": 52.74, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 43.0 }, "checkpoint": "d6c8b863c0d3771a1c6fc4b1ab679f085a4fef9715c91403a832f1be4b0a2d80", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sf20cfcad913002", "s0a314e240c8a61", "s26952945022a2a", "sfa2786fe7555bc", "s9d09e028688960", "s0cc0331c90aa6e", "sa85b49a1125fe0" ] } ], "peak": { "five": { "score": 50.44, "ages": [ 80 ], "rank": 116 }, "six": { "score": 48.05583333333333, "ages": [ 80 ], "rank": 113 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 92, "id": "baby/B_READER_DEPTH8_BATCH4096_BETA2_100M", "original_name": "旧256・8層・単一出力", "title": "256幅・FFN896・8層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 8, "parts": { "width": 256, "ffn": 896, "depth": 8, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "未確認", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 8, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 15, "sources": [ "s935e69ff31fa65", "sd547301a54bbf3", "sbcc344a5a921cd", "sc0efc1a67919cd", "s9eeb0c6a354504", "sd2a432d68c2276", "s01b58e8deb5288", "s711fa171b6d520" ], "points": [ { "age": 80, "six": 48.04, "five": 50.412, "tasks": { "blimp_filtered": 67.94, "supplement_filtered": 59.65, "ewok_filtered": 52.32, "entity_tracking": 19.89, "comps": 52.26, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 51.0 }, "checkpoint": "16a11cd71251564778b7034d04e7ecade528617dd8d3b65d111e3b74f9502624", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "scbced4dceced88", "sbb15bea1f2639c", "s5525b8783fc4d3", "sa5f88999f1a105", "sb648af707a1e43", "s920f977465a5ea", "sb604675e732101", "s87274b802d7080", "sddeb9f5819156f", "se52da469631727", "s278b0c99ba2d67" ] }, { "age": 100, "six": 47.700833333333335, "five": 50.514, "tasks": { "blimp_filtered": 67.5, "supplement_filtered": 60.18, "ewok_filtered": 51.63, "entity_tracking": 21.05, "comps": 52.21, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 43.0 }, "checkpoint": "7e3e027d73871ab2540f9a0b554717fb85ece442ae835df656cf6626ff44d8de", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "scbced4dceced88", "sbb15bea1f2639c", "s105f8748e4423d", "s039cd54062befd", "s920f977465a5ea", "sb604675e732101", "s87274b802d7080", "sddeb9f5819156f", "se52da469631727", "s278b0c99ba2d67" ] } ], "peak": { "five": { "score": 50.514, "ages": [ 100 ], "rank": 108 }, "six": { "score": 48.04, "ages": [ 80 ], "rank": 114 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 93, "id": "run/IPRLM384_A_VARIANCE_INPUT_MINUS5_SEED20260907_40M_20261021", "original_name": "IPRLM384_A_VARIANCE_INPUT_MINUS5_SEED20260907_40M_20261021", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "A＋FFN分散入力だけ−5％", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005066666666666667, 0.0004433333333333333, 0.00037999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All18same-seed40M LRconditions;reuse15controls;allgenres/paired/fullinventory" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 98, "sources": [ "sbdac6f1cf2efa4", "sf2bd87ac7aa3d6" ], "points": [ { "age": 40.0, "six": 48.0325, "five": 50.4, "tasks": { "blimp_filtered": 68.97, "supplement_filtered": 56.99, "ewok_filtered": 53.77, "entity_tracking": 19.69, "comps": 52.58, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 52.0 }, "checkpoint": "10150fe3729d9603aa9f91937f8079f6ba0afc9ea14ea94e298298f9e5c89cf3", "precision": "fp32", "verification": "primary", "sources": [ "scee85629e027ae", "s6649b943d81cf2" ] } ], "peak": { "five": { "score": 50.4, "ages": [ 40.0 ], "rank": 119 }, "six": { "score": 48.0325, "ages": [ 40.0 ], "rank": 115 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 94, "id": "run/IPRLM_PARENT320_BRANCH123_W384_FFN1408_SEED20260907_20261018", "original_name": "IPRLM_PARENT320_BRANCH123_W384_FFN1408_SEED20260907_20261018", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_additive_bypass": false, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at70/80/90/100M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Full known inventory plus matched-age parent320; observed peaks retained separately" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 152, "sources": [ "s6f20c61be1fc2b", "s8861fa9a7f5fb7", "s9d46ffbbaf8e1c" ], "points": [ { "age": 70.0, "six": 48.02583333333333, "five": 49.41, "tasks": { "blimp_filtered": 66.93, "supplement_filtered": 54.77, "ewok_filtered": 52.13, "entity_tracking": 20.9, "comps": 52.32, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 56.0 }, "checkpoint": "9075cdb02242007c8e28d90fb8819cb98c67fb3c06394875f0381b24cdc1cad6", "precision": "fp32", "verification": "primary", "sources": [ "sf58baf6ef1a477", "sa39eebb09d013a" ] }, { "age": 80.0, "six": 47.11833333333333, "five": 49.206, "tasks": { "blimp_filtered": 66.61, "supplement_filtered": 56.04, "ewok_filtered": 51.23, "entity_tracking": 19.92, "comps": 52.23, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 52.0 }, "checkpoint": "967d6a85f79b1a1f0b3e47d6c77f0c87b2c10dddc8e484c5cbf74a07531e93d9", "precision": "fp32", "verification": "primary", "sources": [ "s7edece67bf34c9", "sa39eebb09d013a" ] }, { "age": 90.0, "six": 46.35666666666667, "five": 48.298, "tasks": { "blimp_filtered": 65.51, "supplement_filtered": 53.92, "ewok_filtered": 50.41, "entity_tracking": 20.2, "comps": 51.45, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "85505dd33e069ad5f124b7bbbe78627befb592ef5ba700403b8e6ae945d23a68", "precision": "fp32", "verification": "primary", "sources": [ "sb1d0472be32d24", "sa39eebb09d013a" ] }, { "age": 100.0, "six": 46.24333333333333, "five": 48.762, "tasks": { "blimp_filtered": 66.15, "supplement_filtered": 53.6, "ewok_filtered": 51.79, "entity_tracking": 20.47, "comps": 51.8, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 44.0 }, "checkpoint": "4560675f35dd3363dc3988617826c15c337fab5c26ba4a8de59cd91ddd7b4933", "precision": "fp32", "verification": "primary", "sources": [ "sfe449796e3c3c9", "sa39eebb09d013a" ] } ], "peak": { "five": { "score": 49.41, "ages": [ 70.0 ], "rank": 177 }, "six": { "score": 48.02583333333333, "ages": [ 70.0 ], "rank": 116 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 95, "id": "run/IPRLM320_FFN_G_RMS_SEED731946281_20261014", "original_name": "IPRLM320_FFN_G_RMS_SEED731946281_20261014", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "未確認", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "rms", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_gate_norm": "rms", "ffn_gate_norm_eps": 1e-06 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000 at70/80/90/100M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed parent320 versus FFN g RMS; both seeds; age-matched and observed peaks; no causal proof from two seeds" }, "milestones_words": [ 40000000, 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 81, "sources": [ "s16a6f3ed338933", "s9392718dc25152", "sf4c28399681292" ], "points": [ { "age": 40.0, "six": 48.020833333333336, "five": 49.998, "tasks": { "blimp_filtered": 69.2, "supplement_filtered": 57.97, "ewok_filtered": 50.84, "entity_tracking": 19.72, "comps": 52.26, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 52.0 }, "checkpoint": "f5c8c3e6f464a95e0d6a0a5a84ff961b6f277fe404db555ecf369bd41a142974", "precision": "fp32", "verification": "primary", "sources": [ "s2b49c38bdd5e7b", "s65a5801604bea8" ] } ], "peak": { "five": { "score": 49.998, "ages": [ 40.0 ], "rank": 152 }, "six": { "score": 48.020833333333336, "ages": [ 40.0 ], "rank": 118 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 96, "id": "baby/B_READER_BATCH4096_BETA2_100M", "original_name": "旧256・7層・batch4096＋β₂調整", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "Save checkpoints each10M; score all only after100M; local six axes; no official submission evaluation", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact FP32 eager CUDA graphs; TF32 disabled; choose validated score batching and graph chunk", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768 } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 12, "sources": [ "sc6faec7953c974", "s0e34ef50b9e646", "s9c721b19083b46", "s82371c955c40a5" ], "points": [ { "age": 10, "six": 45.9675, "five": 47.934, "tasks": { "blimp_filtered": 63.72, "supplement_filtered": 55.15, "ewok_filtered": 50.5, "entity_tracking": 19.06, "comps": 51.24, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 48.0 }, "checkpoint": "2a16714f80f10aa10184bc8a81b41756348e0b2a41e03bc6420830cc900bcaff", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s3525c7a8d89c2e", "sce6d1a790673f3", "s383a24f7503c9d", "s46a1f68b097a0c", "sf781c81a7f83ac", "s6232fe567315c0", "s1b2b47d6573dc5" ] }, { "age": 20, "six": 46.47416666666667, "five": 49.13, "tasks": { "blimp_filtered": 65.84, "supplement_filtered": 58.33, "ewok_filtered": 51.17, "entity_tracking": 19.19, "comps": 51.12, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "fe38346bc1a6c6e0e19b1b74046bb71f387d0811be314632b23d00082a6a76d1", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s3525c7a8d89c2e", "sce6d1a790673f3", "s538a800f9eb5c7", "sfdba617f7a2267", "sf781c81a7f83ac", "s6232fe567315c0", "s1b2b47d6573dc5" ] }, { "age": 30, "six": 46.53916666666667, "five": 49.102, "tasks": { "blimp_filtered": 67.27, "supplement_filtered": 55.1, "ewok_filtered": 52.97, "entity_tracking": 18.9, "comps": 51.27, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 49.0 }, "checkpoint": "a53123ca49cdf89c6487bdcd17b640ad5c619fbe8f405d15f38e1f86953ca2f1", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s3525c7a8d89c2e", "sce6d1a790673f3", "s121fe4d4b77923", "s10a6c25e205536", "sf781c81a7f83ac", "s6232fe567315c0", "s1b2b47d6573dc5" ] }, { "age": 40, "six": 47.67666666666667, "five": 50.094, "tasks": { "blimp_filtered": 67.74, "supplement_filtered": 59.73, "ewok_filtered": 52.04, "entity_tracking": 19.22, "comps": 51.74, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 44.0 }, "checkpoint": "32447642a5ff786f8f699d11aef67f44cd758760f1e2d0ec931ab0dfc442a603", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s3525c7a8d89c2e", "sce6d1a790673f3", "s9bbe184c01349b", "s6f190fa13daff4", "sf781c81a7f83ac", "s6232fe567315c0", "s1b2b47d6573dc5" ] }, { "age": 50, "six": 47.200833333333335, "five": 49.608, "tasks": { "blimp_filtered": 67.39, "supplement_filtered": 57.79, "ewok_filtered": 51.21, "entity_tracking": 20.02, "comps": 51.63, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 48.0 }, "checkpoint": "d26e2f9232006c6557e9ec8a91f4aeb3940db3f7e4bbf5729cd3193ac147d401", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s3525c7a8d89c2e", "sce6d1a790673f3", "s895ba99c101375", "sa49e4cd835c899", "sf781c81a7f83ac", "s6232fe567315c0", "s1b2b47d6573dc5" ] }, { "age": 60, "six": 47.263333333333335, "five": 49.774, "tasks": { "blimp_filtered": 66.46, "supplement_filtered": 58.36, "ewok_filtered": 52.9, "entity_tracking": 19.45, "comps": 51.7, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 50.0 }, "checkpoint": "72b6e7e94abc2a323f29b43bdb6e253f8e05baa1e9027a5ab4a18b4dee251106", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s3525c7a8d89c2e", "sce6d1a790673f3", "sd09c26c0d72627", "sdedb5abd189931", "sf781c81a7f83ac", "s6232fe567315c0", "s1b2b47d6573dc5" ] }, { "age": 70, "six": 47.625, "five": 50.32, "tasks": { "blimp_filtered": 67.96, "supplement_filtered": 59.65, "ewok_filtered": 52.69, "entity_tracking": 19.43, "comps": 51.87, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 45.0 }, "checkpoint": "798c11eab37378702180b242adb8d8b49e2cf38cb239d8cf175c10ca5076c28e", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s3525c7a8d89c2e", "sce6d1a790673f3", "sfe38a34447f2ee", "s0b015169f163a9", "sf781c81a7f83ac", "s6232fe567315c0", "s1b2b47d6573dc5" ] }, { "age": 80, "six": 47.995, "five": 50.358, "tasks": { "blimp_filtered": 67.05, "supplement_filtered": 60.54, "ewok_filtered": 52.81, "entity_tracking": 19.38, "comps": 52.01, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 51.0 }, "checkpoint": "6664a161b6cf859ae9814d1e1b590ba047e6ae0479a4a547b28d5e32f4c65bac", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s3525c7a8d89c2e", "sce6d1a790673f3", "sdc41bbf343186a", "sba56100ea7eb4a", "sf781c81a7f83ac", "s6232fe567315c0", "s1b2b47d6573dc5" ] }, { "age": 90, "six": 47.318333333333335, "five": 49.458, "tasks": { "blimp_filtered": 66.72, "supplement_filtered": 57.58, "ewok_filtered": 52.2, "entity_tracking": 18.93, "comps": 51.86, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 48.0 }, "checkpoint": "181a86a7c6fde11b73981e1dd859f031147d89182bcab551c4ad5b248c3498f6", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s3525c7a8d89c2e", "sce6d1a790673f3", "sdaad2239f8641b", "sa6ebaf8ee35503", "sf781c81a7f83ac", "s6232fe567315c0", "s1b2b47d6573dc5" ] }, { "age": 100, "six": 47.51083333333333, "five": 49.592, "tasks": { "blimp_filtered": 68.27, "supplement_filtered": 57.01, "ewok_filtered": 52.08, "entity_tracking": 18.93, "comps": 51.67, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 48.0 }, "checkpoint": "3030c46d2a26eee25ca111d1c4c11dfbc7fe7ddf2cda2bfd7eb4189bdf83ef12", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s3525c7a8d89c2e", "sce6d1a790673f3", "s09a9d1f24fc614", "sfa592be65b39ba", "sf781c81a7f83ac", "s6232fe567315c0", "s1b2b47d6573dc5" ] } ], "peak": { "five": { "score": 50.358, "ages": [ 80 ], "rank": 122 }, "six": { "score": 47.995, "ages": [ 80 ], "rank": 119 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 97, "id": "run/IPRLM320_FFN_CONDITIONAL_BYPASS_SEED731946281_20261015", "original_name": "IPRLM320_FFN_CONDITIONAL_BYPASS_SEED731946281_20261015", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "negative_sigmoid", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true, "ffn_bypass_condition": "negative_sigmoid", "ffn_bypass_threshold": 5.0 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After each40M training; no automatic extension.", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed same40M parent320, unconditional bypass and conditional bypass. 40M is budget allocation, not final peak ranking." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 79, "sources": [ "sb81c91d5737418", "sc6df758ce04a18", "sba01cd3d123504" ], "points": [ { "age": 40.0, "six": 47.16916666666667, "five": 49.77, "tasks": { "blimp_filtered": 68.43, "supplement_filtered": 55.89, "ewok_filtered": 52.47, "entity_tracking": 19.56, "comps": 52.5, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "4e677bfe944242e5978791f0fabc18472f23a48723168cc30349ee3438708472", "precision": "fp32", "verification": "primary", "sources": [ "sad1f4098ea5b92", "sdf42801a85e8be" ] }, { "age": 70.0, "six": 47.945, "five": 50.592, "tasks": { "blimp_filtered": 69.66, "supplement_filtered": 57.5, "ewok_filtered": 52.35, "entity_tracking": 20.79, "comps": 52.66, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 50.0 }, "checkpoint": "70cbb0e9ba470044ec4c9dc4c23a39c2bcf25dde897c7ed5d357fb616501b5f3", "precision": "fp32", "verification": "primary", "sources": [ "se474f82afa3a57", "sa17d6eb3efcbf3" ] } ], "peak": { "five": { "score": 50.592, "ages": [ 70.0 ], "rank": 98 }, "six": { "score": 47.945, "ages": [ 70.0 ], "rank": 120 } }, "clear_points": [ { "number": 97, "age": 70, "seed": 731946281, "checkpoint_sha256": "70cbb0e9ba470044ec4c9dc4c23a39c2bcf25dde897c7ed5d357fb616501b5f3", "old_six": 47.945, "old_five": 50.592, "old_tasks": { "blimp_filtered": 69.66, "supplement_filtered": 57.5, "ewok_filtered": 52.35, "entity_tracking": 20.79, "comps": 52.66, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 50.0 }, "correct": 12234, "total": 17274, "accuracy": 70.82320250086835, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6984, "total": 8392, "ties": 0 }, "comps": { "correct": 3886, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 1, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 13, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 967, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 4, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 4, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 0, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 72, "total": 114, "ties": 0 }, "boolq": { "correct": 9, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 65.18231349900452, "checkpoint_rank": 12, "rank": 11 } ], "clear_peak": { "number": 97, "age": 70, "seed": 731946281, "checkpoint_sha256": "70cbb0e9ba470044ec4c9dc4c23a39c2bcf25dde897c7ed5d357fb616501b5f3", "old_six": 47.945, "old_five": 50.592, "old_tasks": { "blimp_filtered": 69.66, "supplement_filtered": 57.5, "ewok_filtered": 52.35, "entity_tracking": 20.79, "comps": 52.66, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 50.0 }, "correct": 12234, "total": 17274, "accuracy": 70.82320250086835, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6984, "total": 8392, "ties": 0 }, "comps": { "correct": 3886, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 1, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 13, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 967, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 4, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 4, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 0, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 72, "total": 114, "ties": 0 }, "boolq": { "correct": 9, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 65.18231349900452, "checkpoint_rank": 12, "rank": 11 }, "present": true }, { "number": 98, "id": "baby/B_ORIGINAL8_LONG512_HPARAM_100M", "original_name": "修正前8層・添付512設定", "title": "256幅・FFN896・8層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 8, "parts": { "width": 256, "ffn": 896, "depth": 8, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "未確認", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 8, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "Save each10M without stopping training; score80M and100M after100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact FP32 eager CUDA graphs; TF32 disabled; choose validated score batching and graph chunk", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "checkpoint_words": [ 80000000, 100000000 ], "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 68905, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "5167571743565769b6fdc7387ad101402745fffe122104c8e0bb04a5af851ad4", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "405af3d2d6ef1e2825c2e26d919bb21785b47615144941fad986194f66900174", "grouping": "Adjacent original lines, target256 content tokens, max512 except individual long lines, max32 lines; reset at source, visible markers and empty lines", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "context_block_content_tokens_mean": 235.85783324867572, "effective_packed_context_tokens_mean_per_epoch": [ 162.6161508924762, 162.64035644821973, 162.64681248015242, 162.62744592180988, 162.65326902465165, 162.6096972944013, 162.6355147606053, 162.64035644821973, 162.62905962551721, 162.6258322501265 ] }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 8, "sources": [ "s8b6d9ce7489e19", "s3e1620298b5b39", "s2a1adebe482a08", "s2484426bb82c60", "sb59898b48215d9", "sb387997a16a160", "sd73d0a7dfae19d" ], "points": [ { "age": 80, "six": 47.549166666666665, "five": 50.226, "tasks": { "blimp_filtered": 68.67, "supplement_filtered": 58.9, "ewok_filtered": 52.33, "entity_tracking": 19.24, "comps": 51.99, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "68914f4d4c086f9ddfc04e8160d59e66c06e6a183184e4b7444e68ba06dcb32d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s95ebc1a9b2b9ae", "s43fc7fd45425ce", "s3be2e21185959c", "sbe56dbeb415d57", "sdf9520310dcdcb", "sc655c8ac114a47", "seb43a309dcee7a", "sad761c1968ffdf", "sf7c0a8fe2ad0b6" ] }, { "age": 100, "six": 47.935833333333335, "five": 50.59, "tasks": { "blimp_filtered": 69.04, "supplement_filtered": 58.98, "ewok_filtered": 52.63, "entity_tracking": 20.31, "comps": 51.99, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "496a5ed68b85969c065c5009d0d16ab0ca8f14b44dbecb2568d2278bdfb436fd", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s95ebc1a9b2b9ae", "s43fc7fd45425ce", "sf319da8a0e155b", "sa730c1b3ad9704", "sdf9520310dcdcb", "sc655c8ac114a47", "seb43a309dcee7a", "sad761c1968ffdf", "sf7c0a8fe2ad0b6" ] } ], "peak": { "five": { "score": 50.59, "ages": [ 100 ], "rank": 101 }, "six": { "score": 47.935833333333335, "ages": [ 100 ], "rank": 121 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 99, "id": "baby/Q2_READER40_BALANCED", "original_name": "Q2・Reader40 Balanced", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 40, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 40, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 31, "sources": [ "s241975a974674f", "s81d231dd522632", "s11cf5d794be60a", "sdec27714c9e85c", "sa7b1ea3999b2dd" ], "points": [ { "age": 80, "six": 47.740833333333335, "five": 50.15, "tasks": { "blimp_filtered": 68.4, "supplement_filtered": 58.12, "ewok_filtered": 51.81, "entity_tracking": 19.92, "comps": 52.5, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 51.0 }, "checkpoint": "2e9e752be0abef0d668bf8480f930e39a42f9bd6138f438a07c04e43020d7184", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s17c75b0f0337bf", "sc4c4c987c3c1e2", "s7c570bba0d0f98", "s9207856624a340", "sba23a66d494654", "s2c7c776d66e92c", "sddad9a6dd8180d" ] }, { "age": 100, "six": 47.93333333333333, "five": 50.79, "tasks": { "blimp_filtered": 68.86, "supplement_filtered": 59.74, "ewok_filtered": 51.94, "entity_tracking": 20.9, "comps": 52.51, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 44.0 }, "checkpoint": "54041e7675b98351f72087f8a85c6c3f5a4ae02fdc5c7906914eba784aba3f51", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s17c75b0f0337bf", "sc4c4c987c3c1e2", "se9783f70a028b6", "sa19f60d8c8d5db", "sba23a66d494654", "s2c7c776d66e92c", "sddad9a6dd8180d" ] } ], "peak": { "five": { "score": 50.79, "ages": [ 100 ], "rank": 77 }, "six": { "score": 47.93333333333333, "ages": [ 100 ], "rank": 122 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 100, "id": "run/IPRLM320_FFN_G_RMS_SEED20260907_20261014", "original_name": "IPRLM320_FFN_G_RMS_SEED20260907_20261014", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "未確認", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "rms", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_gate_norm": "rms", "ffn_gate_norm_eps": 1e-06 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000 at70/80/90/100M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed parent320 versus FFN g RMS; both seeds; age-matched and observed peaks; no causal proof from two seeds" }, "milestones_words": [ 40000000, 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 81, "sources": [ "s6393487b9f0aca", "s6526d6d09fcf7a", "sbe6f6c30d76788" ], "points": [ { "age": 40.0, "six": 47.92583333333334, "five": 50.272, "tasks": { "blimp_filtered": 69.72, "supplement_filtered": 57.31, "ewok_filtered": 51.9, "entity_tracking": 19.97, "comps": 52.46, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 52.0 }, "checkpoint": "008c689024e603a00a7501beb8e0c222b614bc96402927795aa3061541dd5b36", "precision": "fp32", "verification": "primary", "sources": [ "s5306874c883fae", "sba0405038cdf7b" ] } ], "peak": { "five": { "score": 50.272, "ages": [ 40.0 ], "rank": 126 }, "six": { "score": 47.92583333333334, "ages": [ 40.0 ], "rank": 123 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 101, "id": "run/IPRLM336_FFN1176_ADDITIVE_SEED731946281_20261016", "original_name": "IPRLM336_FFN1176_ADDITIVE_SEED731946281_20261016", "title": "336幅・FFN1176・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 336, "ffn": 1176, "depth": 9, "parts": { "width": 336, "ffn": 1176, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 336, "layers": 9, "hidden": 1176, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed same40M parent320 and additive320; final record49.8675 remains separate" }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 90, "sources": [ "s638ee18480bc28", "s9c1645567d6f89", "sd3e6ee75aee136" ], "points": [ { "age": 40.0, "six": 47.913333333333334, "five": 50.566, "tasks": { "blimp_filtered": 69.04, "supplement_filtered": 58.16, "ewok_filtered": 52.83, "entity_tracking": 20.41, "comps": 52.39, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 46.0 }, "checkpoint": "3f9392221a04b1e5d94bf5ef603d0500b9adc622424e1173a7f7f5df468dfc4a", "precision": "fp32", "verification": "primary", "sources": [ "s6485eb07855fe3", "s8727533a9e5b5e" ] } ], "peak": { "five": { "score": 50.566, "ages": [ 40.0 ], "rank": 104 }, "six": { "score": 47.913333333333334, "ages": [ 40.0 ], "rank": 124 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 102, "id": "run/IPRLM336_FFN1120_ADDITIVE_SEED731946281_20261016", "original_name": "IPRLM336_FFN1120_ADDITIVE_SEED731946281_20261016", "title": "336幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 336, "ffn": 1120, "depth": 9, "parts": { "width": 336, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.3333333333333335 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 336, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M parent320/additive320/additive336FFN1176 and siblingFFN variants; observed later peaks separate." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 89, "sources": [ "sb315b7d15cf158", "s6960502e39e405", "s7bd5609b5ff3da" ], "points": [ { "age": 40.0, "six": 47.87416666666667, "five": 50.41, "tasks": { "blimp_filtered": 69.45, "supplement_filtered": 58.7, "ewok_filtered": 51.19, "entity_tracking": 20.34, "comps": 52.37, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 50.0 }, "checkpoint": "aa470770cd6810d36b1d05ab5eb7495e724d001d4344e39d0cd6c91c51f89cac", "precision": "fp32", "verification": "primary", "sources": [ "s7a23c6045d3578", "sfbf91e93380381" ] } ], "peak": { "five": { "score": 50.41, "ages": [ 40.0 ], "rank": 118 }, "six": { "score": 47.87416666666667, "ages": [ 40.0 ], "rank": 125 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 103, "id": "baby/B_READER_DEPTH9_WIDTH272_FFN960_100M", "original_name": "旧272・9層・FFN960", "title": "272幅・FFN960・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 272, "ffn": 960, "depth": 9, "parts": { "width": 272, "ffn": 960, "depth": 9, "attention": "8×34", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "未確認", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5294117647058822 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 272, "layers": 9, "hidden": 960, "attention_heads": 8, "head_dim": 34, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "history_norm": "separate", "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 25, "sources": [ "s43e98f55c0fdaf", "s11acbfd5154e62", "sba2ba6ef8d6865", "sa3ad90f1fae09f", "s17be4cde261018", "s2b1f98626c8e8d", "sa85cb5d7e26292", "s1d503da12cc2d2" ], "points": [ { "age": 80, "six": 47.845, "five": 50.764, "tasks": { "blimp_filtered": 68.25, "supplement_filtered": 61.01, "ewok_filtered": 52.33, "entity_tracking": 19.82, "comps": 52.41, "global_piqa_parallel": 16.5, "global_piqa_nonparallel": 50.0 }, "checkpoint": "27eebf21ef2d8a736a5fb4b4639487b41b63b85ef8817f51c74c2588b3f6b166", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc2de75a91bacbe", "sf44b27bc1a6260", "sefdb23321f6704", "saebc909f3ff909", "sfbc642f41024d1", "s30fbd4808bb6d7", "scb6121184f53ce", "s9912b4ddaaa548", "sfd6a65764554c7" ] }, { "age": 100, "six": 47.77333333333333, "five": 50.68, "tasks": { "blimp_filtered": 68.98, "supplement_filtered": 58.93, "ewok_filtered": 52.99, "entity_tracking": 19.94, "comps": 52.56, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 49.0 }, "checkpoint": "82bbe1ad3c429ac44a6b4d81fce620b5ebd56f688ee9a81d9e19a7cf99126a8a", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc2de75a91bacbe", "sf44b27bc1a6260", "sf9761dfe86318f", "s41895fcec5f949", "sfbc642f41024d1", "s30fbd4808bb6d7", "scb6121184f53ce", "s9912b4ddaaa548", "sfd6a65764554c7" ] } ], "peak": { "five": { "score": 50.764, "ages": [ 80 ], "rank": 80 }, "six": { "score": 47.845, "ages": [ 80 ], "rank": 126 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 104, "id": "run/IPRLM384_FFN1408_LR10_OTHER90_SEED20260907_40M_20261019", "original_name": "IPRLM384_FFN1408_LR10_OTHER90_SEED20260907_40M_20261019", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "−10% その他97parameterだけ1割減", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005333333333333334, 0.00046666666666666666, 0.00039999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "aftereachcase40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Three LR allocation cases and frozen baseline40M; full inventory; no final ranking from40M alone" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 114, "sources": [ "s57db5ff0fd3870", "s897df5ee6662fc" ], "points": [ { "age": 40.0, "six": 47.810833333333335, "five": 50.14, "tasks": { "blimp_filtered": 69.43, "supplement_filtered": 57.8, "ewok_filtered": 51.46, "entity_tracking": 19.67, "comps": 52.34, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 50.0 }, "checkpoint": "3f93c77ec79885a203e20c92a5e3bff7f8e681b5c1eb1691994dcd3402336c1e", "precision": "fp32", "verification": "primary", "sources": [ "s7d168660523bcb", "s123d3fab8b6cf6" ] } ], "peak": { "five": { "score": 50.14, "ages": [ 40.0 ], "rank": 140 }, "six": { "score": 47.810833333333335, "ages": [ 40.0 ], "rank": 127 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 105, "id": "baby/B_READER_DEPTH9_WIDTH272_ATTN256_READER64_FIXED_100M", "original_name": "旧272・Attention256・Reader64固定", "title": "272幅・FFN960・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 272, "ffn": 960, "depth": 9, "parts": { "width": 272, "ffn": 960, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 64, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "なし", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5294117647058822 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 272, "layers": 9, "hidden": 960, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 64, "history_norm": "separate", "seed": 20260907, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_distance": "none", "learn_rope_base": false, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "candidate_variance_scale_init": 0.1, "attention_width": 256, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score60/70/80/90/100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 20, "sources": [ "s6e984dbfa17552", "s2f61a506d553c8", "s3f98c2ccb9255a", "s393ef734e64837", "s16b87f7f612f5a", "s97c3ea2d079586", "sa7700868b18edc", "s2bee74590aa942" ], "points": [ { "age": 60, "six": 46.86666666666667, "five": 49.498, "tasks": { "blimp_filtered": 66.19, "supplement_filtered": 58.21, "ewok_filtered": 51.65, "entity_tracking": 19.37, "comps": 52.07, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 48.0 }, "checkpoint": "8aef8b66d58363c3a268c4c67e797f2b4db8f9e346bdb2e4d822c1354eae006b", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sbedd7293557f3b", "s16eea0b886506d", "s18963127720ea1", "s3fe1bf4592c2bf", "sf08b61d227d054", "sd24d63c675a470", "scc86239c20ba22", "s994450448aba9a", "scd94c3b666baab" ] }, { "age": 70, "six": 46.903333333333336, "five": 49.836, "tasks": { "blimp_filtered": 69.53, "supplement_filtered": 56.05, "ewok_filtered": 51.43, "entity_tracking": 20.12, "comps": 52.05, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 47.0 }, "checkpoint": "28d08f5a7f4faa0f932bb4e803f940bd544a571ee603af603eb25a6772f49266", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sbedd7293557f3b", "s16eea0b886506d", "s1bd232b3bcc5eb", "s241a489e4ad71f", "sf08b61d227d054", "sd24d63c675a470", "scc86239c20ba22", "s994450448aba9a", "scd94c3b666baab" ] }, { "age": 80, "six": 47.8075, "five": 50.036, "tasks": { "blimp_filtered": 68.61, "supplement_filtered": 58.93, "ewok_filtered": 50.95, "entity_tracking": 19.73, "comps": 51.96, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 51.0 }, "checkpoint": "43baadae1d6ed7128b39a53fea63040a213bf5fa389b257e29d85df946727b59", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sbedd7293557f3b", "s16eea0b886506d", "s13e8b3c991dcb1", "s1eae6a7b0b1f69", "sf08b61d227d054", "sd24d63c675a470", "scc86239c20ba22", "s994450448aba9a", "scd94c3b666baab" ] }, { "age": 90, "six": 46.9375, "five": 49.404, "tasks": { "blimp_filtered": 68.1, "supplement_filtered": 56.57, "ewok_filtered": 50.94, "entity_tracking": 19.77, "comps": 51.64, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 43.0 }, "checkpoint": "a904036eb966a205cac8b2290979e7b814af2521e29f56558cbb7afd8e0a2493", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sbedd7293557f3b", "s16eea0b886506d", "s6aec33ade43b32", "se6da5cea5380db", "sf08b61d227d054", "sd24d63c675a470", "scc86239c20ba22", "s994450448aba9a", "scd94c3b666baab" ] }, { "age": 100, "six": 47.1425, "five": 49.538, "tasks": { "blimp_filtered": 68.52, "supplement_filtered": 55.55, "ewok_filtered": 52.07, "entity_tracking": 19.39, "comps": 52.16, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 48.0 }, "checkpoint": "eab95578bfe867b69cbc4d60fc56632511e571e691ba0b4617bc45626e80aa49", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sbedd7293557f3b", "s16eea0b886506d", "s3c720270faac57", "s26422c1edb1be6", "sf08b61d227d054", "sd24d63c675a470", "scc86239c20ba22", "s994450448aba9a", "scd94c3b666baab" ] } ], "peak": { "five": { "score": 50.036, "ages": [ 80 ], "rank": 147 }, "six": { "score": 47.8075, "ages": [ 80 ], "rank": 128 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 106, "id": "run/IPRLM384_FFN_INPUT_OUTPUT_MATCH_LR_SEED20260907_40M_20261020", "original_name": "IPRLM384_FFN_INPUT_OUTPUT_MATCH_LR_SEED20260907_40M_20261020", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "FFN入口・出口LR一致版", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005333333333333334, 0.00046666666666666666, 0.00039999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All10matched-seed40M LRconditions, allgenres and paired outcomes; fullknowninventory" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 106, "sources": [ "s4fad7799c56d68", "se72fb680edcf33" ], "points": [ { "age": 40.0, "six": 47.80416666666667, "five": 50.038, "tasks": { "blimp_filtered": 68.58, "supplement_filtered": 58.08, "ewok_filtered": 51.4, "entity_tracking": 20.01, "comps": 52.12, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 49.0 }, "checkpoint": "3f44ad2697e9b886945855dfcdac96a99cb0f4e3ce1239bdcdf5d3bdfd31ce71", "precision": "fp32", "verification": "primary", "sources": [ "sc51038aefd04fc", "s1f9aa20b0d5736" ] } ], "peak": { "five": { "score": 50.038, "ages": [ 40.0 ], "rank": 146 }, "six": { "score": 47.80416666666667, "ages": [ 40.0 ], "rank": 129 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 107, "id": "baby/Q1_READER32_BALANCED", "original_name": "Q1・Reader32 Balanced", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 31, "sources": [ "s4a410002f682c2", "s1fa19f1e9c1834", "s6766a55130f308", "sd50480799c648d", "s78876bbd6a8a46" ], "points": [ { "age": 80, "six": 47.711666666666666, "five": 50.512, "tasks": { "blimp_filtered": 68.27, "supplement_filtered": 61.34, "ewok_filtered": 51.26, "entity_tracking": 19.44, "comps": 52.25, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 48.0 }, "checkpoint": "cbcbc62a0d0b471a84be8dd1dd481f273f53a16a5d2fc225217ec131928d4315", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5ae9d341f0fa60", "s8f18b2a6378fee", "sbef9790c3af934", "s83f3988a06df50", "s96a693de8a6ba1", "sc88675e471ca70", "s0aaa9a81ef74cc" ] }, { "age": 100, "six": 47.79833333333333, "five": 50.616, "tasks": { "blimp_filtered": 68.7, "supplement_filtered": 59.27, "ewok_filtered": 52.83, "entity_tracking": 20.07, "comps": 52.21, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 48.0 }, "checkpoint": "9a99254c832a546cad1d2e0d13e94daf6e00d84cba3e00bee08a2b2bea4cc4b2", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5ae9d341f0fa60", "s8f18b2a6378fee", "sa587a90ca2922d", "se880c4b329f6e3", "s96a693de8a6ba1", "sc88675e471ca70", "s0aaa9a81ef74cc" ] } ], "peak": { "five": { "score": 50.616, "ages": [ 100 ], "rank": 95 }, "six": { "score": 47.79833333333333, "ages": [ 100 ], "rank": 130 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 108, "id": "run/IPRLM336_FFN1120_ADDITIVE_SEED20260907_20261016", "original_name": "IPRLM336_FFN1120_ADDITIVE_SEED20260907_20261016", "title": "336幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1120, "depth": 9, "parts": { "width": 336, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.3333333333333335 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 336, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M parent320/additive320/additive336FFN1176 and siblingFFN variants; observed later peaks separate." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 89, "sources": [ "s5fc6eafc458104", "sc529cd61bf9d99", "sf8825af83e0b38" ], "points": [ { "age": 40.0, "six": 47.79666666666667, "five": 49.92, "tasks": { "blimp_filtered": 68.29, "supplement_filtered": 58.02, "ewok_filtered": 50.77, "entity_tracking": 20.59, "comps": 51.93, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 53.0 }, "checkpoint": "7ff4a9a2fdfb6166c4b0ca7dbdbd6bccff4246597df1ad57aa2fe3e765babe9e", "precision": "fp32", "verification": "primary", "sources": [ "sf5fb6392f3dcaa", "se1a3b5bc8aebb4" ] } ], "peak": { "five": { "score": 49.92, "ages": [ 40.0 ], "rank": 156 }, "six": { "score": 47.79666666666667, "ages": [ 40.0 ], "rank": 131 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 109, "id": "run/IPRLM_FFN1121_320_SEED731946281_40M_20261013", "original_name": "IPRLM_FFN1121_320_SEED731946281_40M_20261013", "title": "320幅・FFN1121・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 320, "ffn": 1121, "depth": 9, "parts": { "width": 320, "ffn": 1121, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.503125 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1121, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_preserve_parent_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "After40M training: evaluate40M only", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.715, "comparison_scope": "Same parent320 architecture/LR/data order; one alternate seed; common70/80/90/100M peak and individual tasks. Two runs do not estimate seed variance precisely." }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 155, "sources": [ "s83b7c43e079e36", "sf17ea43370abfa", "s1dfcf32c02bb5c" ], "points": [ { "age": 40.0, "six": 47.759166666666665, "five": 49.984, "tasks": { "blimp_filtered": 68.69, "supplement_filtered": 58.24, "ewok_filtered": 51.41, "entity_tracking": 19.34, "comps": 52.24, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 49.0 }, "checkpoint": "63ac118ff172ab8b46f0b7e4d5b7cfab9684d2c27ee53d07ecabea73daa326b6", "precision": "fp32", "verification": "primary", "sources": [ "sd129bd2e5397ec", "sf67deea19d5778" ] } ], "peak": { "five": { "score": 49.984, "ages": [ 40.0 ], "rank": 154 }, "six": { "score": 47.759166666666665, "ages": [ 40.0 ], "rank": 132 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 110, "id": "run/VISTA336_CANDIDATE_READ_RECEIVED_GATE_READER48_SEED20260907_20261017", "original_name": "VISTA336_CANDIDATE_READ_RECEIVED_GATE_READER48_SEED20260907_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": 48, "received": "あり", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTA_CANDIDATE_READ_RECEIVED_GATE_READER48", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": true, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336, "reader_width": 48, "candidate_variance_scale_init": 0.1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "40M only after firstseed training; no remainingseed auto-launch", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Verified exact recurrent FP32 including original Reader48; same scoring conditions; cached immutable moment table", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M comparison with received-gate baseline, FFNadd child and old336 Reader48 donor; benchmark not run yet." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 163, "sources": [ "s80e98119f2fdfc", "s0ffa0f690593ee", "sdf36bad1586b31" ], "points": [ { "age": 40.0, "six": 47.75333333333333, "five": 50.162, "tasks": { "blimp_filtered": 69.34, "supplement_filtered": 58.13, "ewok_filtered": 50.99, "entity_tracking": 20.01, "comps": 52.34, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 52.0 }, "checkpoint": "c6c0de5a539d0b59a7fe9851fd814156a2a9d2aa7277133a538f89272581e57b", "precision": "fp32", "verification": "primary", "sources": [ "sd535fa5481663f", "s9d4e020d7c046a" ] } ], "peak": { "five": { "score": 50.162, "ages": [ 40.0 ], "rank": 136 }, "six": { "score": 47.75333333333333, "ages": [ 40.0 ], "rank": 133 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 111, "id": "run/IPRLM336_FFN1232_ADDITIVE_SEED731946281_20261016", "original_name": "IPRLM336_FFN1232_ADDITIVE_SEED731946281_20261016", "title": "336幅・FFN1232・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M parent320/additive320/additive336FFN1176 and siblingFFN variants; observed later peaks separate." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 91, "sources": [ "s1e72383b3b69de", "sb15a30d7588733", "scbd3f629687d63" ], "points": [ { "age": 40.0, "six": 47.74916666666667, "five": 50.66, "tasks": { "blimp_filtered": 69.65, "supplement_filtered": 58.26, "ewok_filtered": 52.41, "entity_tracking": 20.31, "comps": 52.67, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "f956b426ac7f422949b66acb03fb356e5293df5eca1f60a4b8871d0d749617db", "precision": "fp32", "verification": "primary", "sources": [ "s4be01eecf4f4e3", "s563741d57c36bf" ] } ], "peak": { "five": { "score": 50.66, "ages": [ 40.0 ], "rank": 87 }, "six": { "score": 47.74916666666667, "ages": [ 40.0 ], "rank": 134 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 112, "id": "baby/B_READER_DEPTH9_WIDTH256_QKNORM_HEAD_GAIN_100M", "original_name": "旧256・9層・QK正規化＋head集中度", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 19, "sources": [ "sa4366f50f69fef", "sdfe48ffe1443ab", "s95f1a450a7fabe", "s0803524cfe59e7", "sabcab643bddf37", "s3e908882e6d510", "s906fd8922b5dac" ], "points": [ { "age": 80, "six": 47.538333333333334, "five": 50.004, "tasks": { "blimp_filtered": 68.03, "supplement_filtered": 58.84, "ewok_filtered": 51.75, "entity_tracking": 18.83, "comps": 52.57, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 51.0 }, "checkpoint": "8374824a69b693b136bdadf96d4d9c40cb86bc1c5fcb1c41392e773ea74f4c89", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s689992eef71c7e", "sf724a7d8974fdf", "sa05044a0a15df6", "s48030250d3f317", "s84e7540bc9b851", "s44db2de204149c", "s3b44cc791f9a52", "s055bedff23e483", "s059065bfd7c4cd" ] }, { "age": 100, "six": 47.714166666666664, "five": 50.424, "tasks": { "blimp_filtered": 68.57, "supplement_filtered": 59.17, "ewok_filtered": 52.75, "entity_tracking": 19.24, "comps": 52.39, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "081352204dda36182a6e25e3e0f527ba40f79fea0af97bf3ad90acce9eec8ffc", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s689992eef71c7e", "sf724a7d8974fdf", "s944b969d0f72a3", "sd73c53465e1bc3", "s84e7540bc9b851", "s44db2de204149c", "s3b44cc791f9a52", "s055bedff23e483", "s059065bfd7c4cd" ] } ], "peak": { "five": { "score": 50.424, "ages": [ 100 ], "rank": 117 }, "six": { "score": 47.714166666666664, "ages": [ 100 ], "rank": 136 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 113, "id": "run/IPRLM384_MAJOR95_OTHER105_MOS100_SEED20260907_40M_20261020", "original_name": "IPRLM384_MAJOR95_OTHER105_MOS100_SEED20260907_40M_20261020", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要−5%・その他＋5%／MoS8行列は元LR", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005066666666666667, 0.0004433333333333333, 0.00037999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All12matchedseed40MLRconditions;reuse11existing;allgenres/pairedoutcomes/fullinventory" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 124, "sources": [ "sa08f2a3ae739d1" ], "points": [ { "age": 40.0, "six": 47.7125, "five": 50.128, "tasks": { "blimp_filtered": 68.4, "supplement_filtered": 58.78, "ewok_filtered": 51.62, "entity_tracking": 19.43, "comps": 52.41, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 47.0 }, "checkpoint": "82dcdc2d6f842a8d0bfbe4f80fbe2d25a4092cb87e65063db3cebd597a28bf4e", "precision": "fp32", "verification": "primary", "sources": [ "sd84f10da8bafab", "s8d6ec34fa07b55" ] } ], "peak": { "five": { "score": 50.128, "ages": [ 40.0 ], "rank": 141 }, "six": { "score": 47.7125, "ages": [ 40.0 ], "rank": 137 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 114, "id": "run/IPRLM_LAYERWISE_320_SEED731946281_40M_20261013", "original_name": "IPRLM_LAYERWISE_320_SEED731946281_40M_20261013", "title": "320幅・FFN1120・9層", "topology": "MoS3・各層分離", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8層 / 7層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "各層を個別", "reference_layers": [ [ 9 ], [ 8 ], [ 7 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_layerwise987", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8 ], [ 7 ] ], "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "After40M training: evaluate40M only", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.715, "comparison_scope": "Same parent320 architecture/LR/data order; one alternate seed; common70/80/90/100M peak and individual tasks. Two runs do not estimate seed variance precisely." }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 155, "sources": [ "sa4234ad92edf6b", "s9befbd17b4b888", "sf08dd27c585573" ], "points": [ { "age": 40.0, "six": 47.67333333333333, "five": 49.478, "tasks": { "blimp_filtered": 68.98, "supplement_filtered": 55.85, "ewok_filtered": 51.25, "entity_tracking": 18.92, "comps": 52.39, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 54.0 }, "checkpoint": "1b4d652a3b505716e9826ca697c00adeed89fb11add924464b189e164e756258", "precision": "fp32", "verification": "primary", "sources": [ "se206df24c94444", "s96fe0e444a29e3" ] } ], "peak": { "five": { "score": 49.478, "ages": [ 40.0 ], "rank": 175 }, "six": { "score": 47.67333333333333, "ages": [ 40.0 ], "rank": 138 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 115, "id": "run/IPRLM384_MAJOR90_OTHER_ROLE_JOINT_SEED20260907_40M_20261021", "original_name": "IPRLM384_MAJOR90_OTHER_ROLE_JOINT_SEED20260907_40M_20261021", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要−10％＋部品別配合（旧失敗配合）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All19same-seed40M LRconditions;all7tasks,5/6aggregates,paired outcomes;fullinventory" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 122, "sources": [ "scbd459cbd8eae7", "s4346e892ca7388" ], "points": [ { "age": 40.0, "six": 47.630833333333335, "five": 49.73, "tasks": { "blimp_filtered": 67.31, "supplement_filtered": 57.57, "ewok_filtered": 51.62, "entity_tracking": 20.0, "comps": 52.15, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 50.0 }, "checkpoint": "70f7f67e998c28ab024add598d76543cee3fb5282908bb019f1d42cf5da88d4d", "precision": "fp32", "verification": "primary", "sources": [ "s0c5a35e98884e3", "s1a18ea05fe4251" ] } ], "peak": { "five": { "score": 49.73, "ages": [ 40.0 ], "rank": 163 }, "six": { "score": 47.630833333333335, "ages": [ 40.0 ], "rank": 139 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 116, "id": "run/IPRLM384_FFN1408_LRPLUS10_OTHER110_SEED20260907_40M_20261020", "original_name": "IPRLM384_FFN1408_LRPLUS10_OTHER110_SEED20260907_40M_20261020", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "その他97parameterだけ1割増", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005333333333333334, 0.00046666666666666666, 0.00039999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "aftereachcase40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Three LR allocation cases and frozen baseline40M; full inventory; no final ranking from40M alone" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 117, "sources": [ "s46d2638435625e", "sd2c9488531ee47" ], "points": [ { "age": 40.0, "six": 47.605, "five": 50.002, "tasks": { "blimp_filtered": 68.56, "supplement_filtered": 56.63, "ewok_filtered": 52.45, "entity_tracking": 19.57, "comps": 52.8, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 46.0 }, "checkpoint": "3c0583b25c740b8991f1152b5a704ffb4b98531b94dd438135606eb0db79fe20", "precision": "fp32", "verification": "primary", "sources": [ "sb7a743355309bf", "s90dd3b917043d9" ] } ], "peak": { "five": { "score": 50.002, "ages": [ 40.0 ], "rank": 150 }, "six": { "score": 47.605, "ages": [ 40.0 ], "rank": 140 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 117, "id": "baby/B_READER_DEPTH9_WIDTH272_ATTN256_READER64_LEARNED_LEN256_LR10_100M", "original_name": "旧272・Reader64可変・長さ256＋10段LR", "title": "272幅・FFN960・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 272, "ffn": 960, "depth": 9, "parts": { "width": 272, "ffn": 960, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 64, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5294117647058822 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 272, "layers": 9, "hidden": 960, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 64, "history_norm": "separate", "seed": 20260907, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "attention_width": 256, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score60/70/80/90/100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "da8e5e14862589add823740a5f5f3a5eedf6a36ac881e6715ab967de87315430", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "packing_change": "Fixed256 row packing only; context groups and each epoch nonpad token stream exactly preserved" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 22, "sources": [ "s1bf013ed2cbab8", "s00b56425dd5ed9", "s6264e6f59ff660", "s9046fa3be735cc", "sb8bebb2fcc3d22", "se81694a9b37579", "scb2c51e5f2fb5c" ], "points": [ { "age": 60, "six": 46.77583333333333, "five": 49.304, "tasks": { "blimp_filtered": 66.16, "supplement_filtered": 56.54, "ewok_filtered": 52.44, "entity_tracking": 19.37, "comps": 52.01, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 44.0 }, "checkpoint": "5a6b1704742e39372af293b5c7d41d43c6c343e92a553d614bfcb5cdfb2c004e", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s53ffb8305a1fdd", "sc95e736925f602", "s7e1f6ab3435cf5", "s4d7d0bc6a33b31", "s7e340fd4ad2b2c", "sb8dfdc88bd38bb", "s24b86a78ef79fd", "s3c2ae79fe6ade6", "sa26e9d649c568f" ] }, { "age": 70, "six": 46.7975, "five": 49.23, "tasks": { "blimp_filtered": 66.09, "supplement_filtered": 55.48, "ewok_filtered": 53.35, "entity_tracking": 19.19, "comps": 52.04, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 45.0 }, "checkpoint": "91fb3981ec2d353b76267a5874a4deab0ed55257794b28405da9faf118e6780b", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s53ffb8305a1fdd", "sc95e736925f602", "s0a01d9bd3159e6", "s613fe3fed5e202", "s7e340fd4ad2b2c", "sb8dfdc88bd38bb", "s24b86a78ef79fd", "s3c2ae79fe6ade6", "sa26e9d649c568f" ] }, { "age": 80, "six": 47.586666666666666, "five": 49.856, "tasks": { "blimp_filtered": 65.77, "supplement_filtered": 59.06, "ewok_filtered": 52.76, "entity_tracking": 19.81, "comps": 51.88, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 55.0 }, "checkpoint": "ea15ac38f0629c4e6eb1c634cc09b22a36438d09966e6a3eeeeb986af9da5309", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s53ffb8305a1fdd", "sc95e736925f602", "sf6641602559b5f", "s30b2f0e685d546", "s7e340fd4ad2b2c", "sb8dfdc88bd38bb", "s24b86a78ef79fd", "s3c2ae79fe6ade6", "sa26e9d649c568f" ] }, { "age": 90, "six": 47.0, "five": 48.576, "tasks": { "blimp_filtered": 65.35, "supplement_filtered": 56.05, "ewok_filtered": 50.33, "entity_tracking": 19.62, "comps": 51.53, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 53.0 }, "checkpoint": "43bd656d6ebb1323c0b349a8b4fb8391486bf39433d5a7c074d92baea02ed67c", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s53ffb8305a1fdd", "sc95e736925f602", "sca0c98152f87a6", "s5330ce662aaa42", "s7e340fd4ad2b2c", "sb8dfdc88bd38bb", "s24b86a78ef79fd", "s3c2ae79fe6ade6", "sa26e9d649c568f" ] }, { "age": 100, "six": 47.46083333333333, "five": 49.726, "tasks": { "blimp_filtered": 67.28, "supplement_filtered": 58.75, "ewok_filtered": 51.64, "entity_tracking": 19.3, "comps": 51.66, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 48.0 }, "checkpoint": "6f50c6a5dd5490341c0073b717a284f3614f0e132bf18f7f2d50c4fb3b5d2811", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s53ffb8305a1fdd", "sc95e736925f602", "s4f5d5391e8b6e1", "sd51a9c6120b3c8", "s7e340fd4ad2b2c", "sb8dfdc88bd38bb", "s24b86a78ef79fd", "s3c2ae79fe6ade6", "sa26e9d649c568f" ] } ], "peak": { "five": { "score": 49.856, "ages": [ 80 ], "rank": 158 }, "six": { "score": 47.586666666666666, "ages": [ 80 ], "rank": 142 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 118, "id": "run/IPRLM384_A_READER_TRANSFORM_PLUS5_SEED20260907_40M_20261021", "original_name": "IPRLM384_A_READER_TRANSFORM_PLUS5_SEED20260907_40M_20261021", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "A＋Candidate Reader変換だけ＋5％", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005066666666666667, 0.0004433333333333333, 0.00037999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All18same-seed40M LRconditions;reuse15controls;allgenres/paired/fullinventory" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 97, "sources": [ "saed3672321c4b6", "s9a9a9f50ca9b46" ], "points": [ { "age": 40.0, "six": 47.58166666666666, "five": 50.068, "tasks": { "blimp_filtered": 68.91, "supplement_filtered": 58.18, "ewok_filtered": 51.47, "entity_tracking": 19.5, "comps": 52.28, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 47.0 }, "checkpoint": "05a2df400fc9406bb9b571a550c3a49f00c3dafe8c00c264a74e1d74e0e5d538", "precision": "fp32", "verification": "primary", "sources": [ "s37d23b0f997c2c", "sed65cc1123fd33" ] } ], "peak": { "five": { "score": 50.068, "ages": [ 40.0 ], "rank": 145 }, "six": { "score": 47.58166666666666, "ages": [ 40.0 ], "rank": 143 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 119, "id": "baby/B_READER_BATCH4096_100M", "original_name": "旧Reader・batch4096", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "Save checkpoints each10M; score all only after100M; local six axes; no official submission evaluation", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 11, "sources": [ "s3eeac0c985e251", "s7231259d886a1d", "sd2395bcf1826e1", "s35e0ac4c16f157" ], "points": [ { "age": 10, "six": 45.69166666666667, "five": 47.7, "tasks": { "blimp_filtered": 63.91, "supplement_filtered": 56.53, "ewok_filtered": 49.92, "entity_tracking": 17.77, "comps": 50.37, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 48.0 }, "checkpoint": "e1f0992a9910abb5a0e70dfbddea51a035052e71451fc14691329dcb6084d2ab", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s0ddf685f760cdc", "s45c1538cc46554", "sc18dc8d214ac78", "s3fefe762921c9a", "sb6a61b65b7d3d6", "s10c6cb6ac8117b", "s3bbb82dfa62b3a" ] }, { "age": 20, "six": 46.13916666666667, "five": 48.522, "tasks": { "blimp_filtered": 65.43, "supplement_filtered": 55.98, "ewok_filtered": 52.44, "entity_tracking": 18.95, "comps": 49.81, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 50.0 }, "checkpoint": "ac5c443e32634129bf23c026df7b252f2834764c11d247e2af2853a9a2cb76e3", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s0ddf685f760cdc", "s45c1538cc46554", "s21d20ed8f14b85", "s6b4848b09420a5", "sb6a61b65b7d3d6", "s10c6cb6ac8117b", "s3bbb82dfa62b3a" ] }, { "age": 30, "six": 46.528333333333336, "five": 48.804, "tasks": { "blimp_filtered": 66.22, "supplement_filtered": 57.89, "ewok_filtered": 51.78, "entity_tracking": 17.96, "comps": 50.17, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 47.0 }, "checkpoint": "f41cd4e182a2716499fe209feef77fd1b15a9f755338048b584232e38ec9ec8c", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s0ddf685f760cdc", "s45c1538cc46554", "s1675f47383af38", "s26ee269089afd7", "sb6a61b65b7d3d6", "s10c6cb6ac8117b", "s3bbb82dfa62b3a" ] }, { "age": 40, "six": 47.26166666666666, "five": 49.578, "tasks": { "blimp_filtered": 67.58, "supplement_filtered": 57.15, "ewok_filtered": 52.74, "entity_tracking": 19.25, "comps": 51.17, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 50.0 }, "checkpoint": "6cacaf1f4375872484c4d24eb7c097b036107eaa2bada77eccb0e44184f3979a", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s0ddf685f760cdc", "s45c1538cc46554", "s97879cdd908459", "s94e628981e209b", "sb6a61b65b7d3d6", "s10c6cb6ac8117b", "s3bbb82dfa62b3a" ] }, { "age": 50, "six": 47.56166666666667, "five": 49.838, "tasks": { "blimp_filtered": 67.02, "supplement_filtered": 60.11, "ewok_filtered": 51.17, "entity_tracking": 19.98, "comps": 50.91, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 51.0 }, "checkpoint": "5a17a6e45fbc8b9e9b01e58836d608cfa8700cb4a0a164594afb922e4ab1ad1c", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s0ddf685f760cdc", "s45c1538cc46554", "s1c6cd870c67ec4", "s80e3b9f9075e55", "sb6a61b65b7d3d6", "s10c6cb6ac8117b", "s3bbb82dfa62b3a" ] }, { "age": 60, "six": 47.041666666666664, "five": 49.314, "tasks": { "blimp_filtered": 66.65, "supplement_filtered": 58.18, "ewok_filtered": 51.11, "entity_tracking": 19.77, "comps": 50.86, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 50.0 }, "checkpoint": "37f5312ced1f2f372a707347d94ad15f34b135d5251bda39506019110dae5c91", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s0ddf685f760cdc", "s45c1538cc46554", "s9e82a65d7726c0", "s9b336d3b0583e8", "sb6a61b65b7d3d6", "s10c6cb6ac8117b", "s3bbb82dfa62b3a" ] }, { "age": 70, "six": 47.195, "five": 50.192, "tasks": { "blimp_filtered": 68.44, "supplement_filtered": 60.13, "ewok_filtered": 50.69, "entity_tracking": 20.49, "comps": 51.21, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 45.0 }, "checkpoint": "23a17c5f1d48c09b76486d5d61b1d5bfbf836695661d090ee275c1ae537c57e0", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s0ddf685f760cdc", "s45c1538cc46554", "s4ca75de2b78ada", "s31ba39dc2b7236", "sb6a61b65b7d3d6", "s10c6cb6ac8117b", "s3bbb82dfa62b3a" ] }, { "age": 80, "six": 46.74666666666667, "five": 49.84, "tasks": { "blimp_filtered": 66.37, "supplement_filtered": 60.52, "ewok_filtered": 51.12, "entity_tracking": 19.81, "comps": 51.38, "global_piqa_parallel": 14.56, "global_piqa_nonparallel": 48.0 }, "checkpoint": "f2b2be4fab8716d1d4fc3c37676c365bca258ef41ea3e5704ba9749fc0401ceb", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s0ddf685f760cdc", "s45c1538cc46554", "s1da76f0b8a38d2", "s535392d7b352b3", "sb6a61b65b7d3d6", "s10c6cb6ac8117b", "s3bbb82dfa62b3a" ] }, { "age": 90, "six": 47.32666666666667, "five": 49.562, "tasks": { "blimp_filtered": 66.29, "supplement_filtered": 58.54, "ewok_filtered": 51.48, "entity_tracking": 20.25, "comps": 51.25, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 49.0 }, "checkpoint": "8ce441097eeb50ee585bb2b54aa41c06e0e481881392a7c5020b7fd626be775d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s0ddf685f760cdc", "s45c1538cc46554", "s4405387e359699", "s971e6d4ac17f34", "sb6a61b65b7d3d6", "s10c6cb6ac8117b", "s3bbb82dfa62b3a" ] }, { "age": 100, "six": 47.48083333333334, "five": 49.844, "tasks": { "blimp_filtered": 66.89, "supplement_filtered": 60.25, "ewok_filtered": 50.57, "entity_tracking": 19.68, "comps": 51.83, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 49.0 }, "checkpoint": "904ebbe7787d76d2f3f62a5f8e9a04094ad43f5060014746206acb97b4333a0c", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s0ddf685f760cdc", "s45c1538cc46554", "s5eab781c54b903", "sdac85f601dcedb", "sb6a61b65b7d3d6", "s10c6cb6ac8117b", "s3bbb82dfa62b3a" ] } ], "peak": { "five": { "score": 50.192, "ages": [ 70 ], "rank": 134 }, "six": { "score": 47.56166666666667, "ages": [ 50 ], "rank": 144 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 120, "id": "run/IPRLM384_FFN1408_LR10_ALL90_SEED20260907_40M_20261019", "original_name": "IPRLM384_FFN1408_LR10_ALL90_SEED20260907_40M_20261019", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "−10% 全LRを1割減", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00042, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "aftereachcase40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Three LR allocation cases and frozen baseline40M; full inventory; no final ranking from40M alone" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 112, "sources": [ "sb4f8c0eb152983", "s4a2590783390bd" ], "points": [ { "age": 40.0, "six": 47.505, "five": 50.27, "tasks": { "blimp_filtered": 69.78, "supplement_filtered": 57.81, "ewok_filtered": 51.34, "entity_tracking": 19.7, "comps": 52.72, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 46.0 }, "checkpoint": "ef7a3e202ea5a19c1ff1693806d4deebd1dd0c90d4f817edb7f473ccd07118a2", "precision": "fp32", "verification": "primary", "sources": [ "s23c3a6807bcbaa", "s22cd02730a3c9d" ] } ], "peak": { "five": { "score": 50.27, "ages": [ 40.0 ], "rank": 128 }, "six": { "score": 47.505, "ages": [ 40.0 ], "rank": 145 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 121, "id": "run/VISTA336_CANDIDATE_READ_RECEIVED_GATE_SEED20260907_20261017", "original_name": "VISTA336_CANDIDATE_READ_RECEIVED_GATE_SEED20260907_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": "未確認", "received": "あり", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTA_CANDIDATE_READ_RECEIVED_GATE", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": true, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "40M only after firstseed training; no remainingseed auto-launch", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 with bounded live graph/KV cache; validated native gate included", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus current candidate-read baseline; only Attention received gate added." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 159, "sources": [ "s553ea5994137bc", "s9f1abde72c7022", "s9d9de68d818646" ], "points": [ { "age": 40.0, "six": 47.44583333333333, "five": 50.208, "tasks": { "blimp_filtered": 69.5, "supplement_filtered": 56.92, "ewok_filtered": 52.04, "entity_tracking": 20.42, "comps": 52.16, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 43.0 }, "checkpoint": "f77f3e6fce363d29b45d7b13ccbec05bc5b0be734316642fa6cb489b60d9424a", "precision": "fp32", "verification": "primary", "sources": [ "s62809d01259c14", "s6ecf3a8763bf16" ] } ], "peak": { "five": { "score": 50.208, "ages": [ 40.0 ], "rank": 133 }, "six": { "score": 47.44583333333333, "ages": [ 40.0 ], "rank": 146 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 122, "id": "baby/B_READER_BATCH8192_80M", "original_name": "旧Reader・batch8192", "title": "256幅・FFN896・7層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "未確認", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluations; preserve10M checkpoints; score80M only after training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 13, "sources": [ "s081197d41160fe", "s435766b2470c61", "sc68321427ff810", "s155cbfafb509b6", "s8f8ee8e7184304", "s766c4b748e8bca", "sc2a221f36478cd" ], "points": [ { "age": 80, "six": 47.43333333333333, "five": 49.99, "tasks": { "blimp_filtered": 68.43, "supplement_filtered": 59.3, "ewok_filtered": 51.55, "entity_tracking": 18.66, "comps": 52.01, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 46.0 }, "checkpoint": "98ec33400cff7a78749f0800d743a0939412b1a20ff9617cb7fa4d2e17097138", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sbc758873358b3c", "sa9fcd6ac0cbbca", "s795a25fb465be1", "s89e7fa65e3356b", "s4d8245a2d3f4e7", "se248ca3e842ada", "sc00d906d7e595c", "scc27fea1434f49", "sc9f9cbce6973bb" ] } ], "peak": { "five": { "score": 49.99, "ages": [ 80 ], "rank": 153 }, "six": { "score": 47.43333333333333, "ages": [ 80 ], "rank": 148 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 123, "id": "run/IPRLM336_FFN1176_ADDITIVE_SEED137_20261016", "original_name": "IPRLM336_FFN1176_ADDITIVE_SEED137_20261016", "title": "336幅・FFN1176・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 336, "ffn": 1176, "depth": 9, "parts": { "width": 336, "ffn": 1176, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 336, "layers": 9, "hidden": 1176, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed same40M parent320 and additive320; final record49.8675 remains separate" }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 90, "sources": [ "sf0cdac41e8df4a", "s80ce697b9cb484", "s346de181f3daf1" ], "points": [ { "age": 40.0, "six": 47.43083333333333, "five": 49.684, "tasks": { "blimp_filtered": 68.57, "supplement_filtered": 56.32, "ewok_filtered": 52.22, "entity_tracking": 19.7, "comps": 51.61, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 50.0 }, "checkpoint": "eb83ea25158934c603283f1a3e0ab8897387862975286b3625277b6179da6422", "precision": "fp32", "verification": "primary", "sources": [ "s687820b26cf7e9", "s1edbc5f89c25a3" ] } ], "peak": { "five": { "score": 49.684, "ages": [ 40.0 ], "rank": 166 }, "six": { "score": 47.43083333333333, "ages": [ 40.0 ], "rank": 149 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 124, "id": "run/IPRLM384_LR_C_MAJOR925_OTHER1025_SEED20260907_40M_20261020", "original_name": "IPRLM384_LR_C_MAJOR925_OTHER1025_SEED20260907_40M_20261020", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要−7.5%・その他＋2.5%", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0004933333333333334, 0.0004316666666666667, 0.00037 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All15matchedseed40MLRconditions;reuse12controls;allgenres/pairedoutcomes/fullinventory" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 120, "sources": [ "s71c508f291d536" ], "points": [ { "age": 40.0, "six": 47.41916666666667, "five": 50.264, "tasks": { "blimp_filtered": 68.59, "supplement_filtered": 59.69, "ewok_filtered": 51.16, "entity_tracking": 19.81, "comps": 52.07, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "3a9eeaae3dfa8b4894826638d67d99345e1623a010343919794b08cc8055709f", "precision": "fp32", "verification": "primary", "sources": [ "s82708912c61bef", "s607a5eca6d8604" ] } ], "peak": { "five": { "score": 50.264, "ages": [ 40.0 ], "rank": 129 }, "six": { "score": 47.41916666666667, "ages": [ 40.0 ], "rank": 150 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 125, "id": "baby/B_DIRECT_BATCH4096_BETA2_80M", "original_name": "直接帰還・batch4096＋β₂調整", "title": "256幅・FFN896・7層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": "未確認", "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "未確認", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluations; preserve10M checkpoints; score80M only after training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 5, "sources": [ "scd8ca61bab3ee4", "sb0b82b94cb04b4", "s589bf2b11a3ba3", "s18f1678e929804", "se5833bde3c7bba", "s3e12d1e16922e0", "sb3877176e5d17b" ], "points": [ { "age": 80, "six": 47.40833333333333, "five": 50.148, "tasks": { "blimp_filtered": 67.17, "supplement_filtered": 60.04, "ewok_filtered": 51.81, "entity_tracking": 20.04, "comps": 51.68, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 48.0 }, "checkpoint": "d9a3aebae04194103c1bd01676ef193a3466d36b4be4fadbe4881e4b4bdf4973", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5373791001d61a", "s41dc53109a6d03", "s17aee91f061f1e", "sd7777724659a19", "sac0c5f853c3aa4", "s95f05f2fa77d8a", "s61abf00776b7ab", "sfd33246765879e", "s1a5689f2710b29" ] } ], "peak": { "five": { "score": 50.148, "ages": [ 80 ], "rank": 137 }, "six": { "score": 47.40833333333333, "ages": [ 80 ], "rank": 151 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 126, "id": "run/IPRLM352_FFN1232_ADDITIVE_SEED731946281_20261016", "original_name": "IPRLM352_FFN1232_ADDITIVE_SEED731946281_20261016", "title": "352幅・FFN1232・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 352, "ffn": 1232, "depth": 9, "parts": { "width": 352, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 352, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M parent320/additive320/additive336; broader peak record separate" }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 95, "sources": [ "sce239fc2e5e57d", "s4f7f072ad1a2d7", "s4301e5b806d231" ], "points": [ { "age": 40.0, "six": 47.395, "five": 50.238, "tasks": { "blimp_filtered": 67.67, "supplement_filtered": 58.77, "ewok_filtered": 51.73, "entity_tracking": 20.33, "comps": 52.69, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 45.0 }, "checkpoint": "fe29c1230f17211d84de3bb024c193c5101c7b1f4135308b0b73145e8c017ec1", "precision": "fp32", "verification": "primary", "sources": [ "sbca3e277db88a5", "s036d531ea6bda1" ] } ], "peak": { "five": { "score": 50.238, "ages": [ 40.0 ], "rank": 132 }, "six": { "score": 47.395, "ages": [ 40.0 ], "rank": 152 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 127, "id": "run/IPRLM384_NORM820_SPREAD_MINUS5_SEED20260907_40M_20261022", "original_name": "IPRLM384_NORM820_SPREAD_MINUS5_SEED20260907_40M_20261022", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "Norm820元版＋FFN分散入力LR−5％", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All22same-seed40Mconditions;all7tasksandpairedoutcomes;fullknowninventory;deltasagainstoriginalNorm820" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 138, "sources": [ "sac18eb400fa0a9", "s3395fc9a95c407" ], "points": [ { "age": 40.0, "six": 47.3925, "five": 50.032, "tasks": { "blimp_filtered": 68.3, "supplement_filtered": 57.61, "ewok_filtered": 51.77, "entity_tracking": 20.25, "comps": 52.23, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 48.0 }, "checkpoint": "3849d52854b3a2e54b685e2d501483c46f38d032e69e50060832f39bc849b377", "precision": "fp32", "verification": "primary", "sources": [ "sedd3dfae55d26b", "sdb3db83171f631" ] } ], "peak": { "five": { "score": 50.032, "ages": [ 40.0 ], "rank": 148 }, "six": { "score": 47.3925, "ages": [ 40.0 ], "rank": 153 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 128, "id": "run/IPRLM352_FFN1232_ADDITIVE_SEED137_20261016", "original_name": "IPRLM352_FFN1232_ADDITIVE_SEED137_20261016", "title": "352幅・FFN1232・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 352, "ffn": 1232, "depth": 9, "parts": { "width": 352, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 352, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M parent320/additive320/additive336; broader peak record separate" }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 95, "sources": [ "s35c34a9532a6c5", "s088717b740de87", "s59fe5fc314b860" ], "points": [ { "age": 40.0, "six": 47.36083333333333, "five": 50.0, "tasks": { "blimp_filtered": 68.74, "supplement_filtered": 56.63, "ewok_filtered": 52.72, "entity_tracking": 19.89, "comps": 52.02, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "f16f4ef4891ffb0b3c5ad81c4f50e01cb0be2e1c25b70e6e006d3de14afea5f9", "precision": "fp32", "verification": "primary", "sources": [ "sc521aeed7057bf", "s527cc0c895ba35" ] } ], "peak": { "five": { "score": 50.0, "ages": [ 40.0 ], "rank": 151 }, "six": { "score": 47.36083333333333, "ages": [ 40.0 ], "rank": 154 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 129, "id": "baby/B_CANDIDATE_READER_100M", "original_name": "Candidate Reader初版", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "Save checkpoints each10M; score all only after100M; local six axes; no official submission evaluation", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 0, "sources": [ "sfb72008b98840c", "s0169eb51b3e20e", "s7c008743fd7e58", "sd80bdf7cb8f907" ], "points": [ { "age": 10, "six": 45.23583333333333, "five": 47.938, "tasks": { "blimp_filtered": 64.43, "supplement_filtered": 56.03, "ewok_filtered": 50.76, "entity_tracking": 18.2, "comps": 50.27, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 45.0 }, "checkpoint": "bcbb41acb20dc6f95a6c567e7dcb53739bcf1e3e7e860868be662db395facc08", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s50eb37b351b870", "sd162e9a04349f7", "sc66a9400a9359c", "s469197596c628c", "s61509c2e16a7cf", "sb8540f9fdc89e7", "sd702222d1cd47e" ] }, { "age": 20, "six": 45.943333333333335, "five": 48.69, "tasks": { "blimp_filtered": 66.19, "supplement_filtered": 55.37, "ewok_filtered": 51.67, "entity_tracking": 18.4, "comps": 51.82, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 45.0 }, "checkpoint": "4c84a48dcb4515d0c4150eea0ec78afad64a541264d038b66c21fab47ad499ae", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s50eb37b351b870", "sd162e9a04349f7", "s6df2d05a22fea5", "s9d9bfc248b5a90", "s61509c2e16a7cf", "sb8540f9fdc89e7", "sd702222d1cd47e" ] }, { "age": 30, "six": 46.45916666666667, "five": 49.012, "tasks": { "blimp_filtered": 66.03, "supplement_filtered": 57.84, "ewok_filtered": 51.17, "entity_tracking": 18.58, "comps": 51.44, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "04a5f4abd4d9752159109906e9e971ca99468a50f759b86be9322456a2453956", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s50eb37b351b870", "sd162e9a04349f7", "s9c5fe913edbe2e", "s9751c9a323947b", "s61509c2e16a7cf", "sb8540f9fdc89e7", "sd702222d1cd47e" ] }, { "age": 40, "six": 46.839166666666664, "five": 49.374, "tasks": { "blimp_filtered": 67.16, "supplement_filtered": 56.21, "ewok_filtered": 52.39, "entity_tracking": 18.84, "comps": 52.27, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "bd4073a40acba8e2a683df0d8cbaeba62a8788165a4100bf60bf06a6b622eb01", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s50eb37b351b870", "sd162e9a04349f7", "s729507d63e3e8b", "s450f6849cfb68d", "s61509c2e16a7cf", "sb8540f9fdc89e7", "sd702222d1cd47e" ] }, { "age": 50, "six": 46.77916666666667, "five": 49.49, "tasks": { "blimp_filtered": 67.1, "supplement_filtered": 58.22, "ewok_filtered": 50.69, "entity_tracking": 19.22, "comps": 52.22, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 48.0 }, "checkpoint": "03b34fc81deb84a1247aa8fa89f9ef158431e185d74a37d3c8d72043a9b8eb13", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s50eb37b351b870", "sd162e9a04349f7", "s111c6d9811b0ae", "s3d5c846d8898ad", "s61509c2e16a7cf", "sb8540f9fdc89e7", "sd702222d1cd47e" ] }, { "age": 60, "six": 46.67333333333333, "five": 49.278, "tasks": { "blimp_filtered": 67.5, "supplement_filtered": 57.49, "ewok_filtered": 50.44, "entity_tracking": 18.67, "comps": 52.29, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 44.0 }, "checkpoint": "41d550c035ff138d6b2c708d251a0aa02d1942c416bf41a24b54a48b0a13fbe8", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s50eb37b351b870", "sd162e9a04349f7", "sb47db0c0d831bd", "s8d3f940ed772db", "s61509c2e16a7cf", "sb8540f9fdc89e7", "sd702222d1cd47e" ] }, { "age": 70, "six": 46.776666666666664, "five": 49.584, "tasks": { "blimp_filtered": 67.04, "supplement_filtered": 59.03, "ewok_filtered": 50.9, "entity_tracking": 18.83, "comps": 52.12, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 48.0 }, "checkpoint": "dff3524c77acb594fd8c238f5b6932c347253d24e0a78edba87d92be06acb41e", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s50eb37b351b870", "sd162e9a04349f7", "s17606e04038dc8", "sec23acc5ab3ab1", "s61509c2e16a7cf", "sb8540f9fdc89e7", "sd702222d1cd47e" ] }, { "age": 80, "six": 47.2725, "five": 49.494, "tasks": { "blimp_filtered": 67.62, "supplement_filtered": 57.9, "ewok_filtered": 50.42, "entity_tracking": 19.43, "comps": 52.1, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 50.0 }, "checkpoint": "3a0e91eaa0f608cd0e4b32410e5499ef72769341c209a2ffc8e0477f48ead4ee", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s50eb37b351b870", "sd162e9a04349f7", "sb0ad566b9db1f3", "s69c81598f72d5a", "s61509c2e16a7cf", "sb8540f9fdc89e7", "sd702222d1cd47e" ] }, { "age": 90, "six": 47.3575, "five": 49.408, "tasks": { "blimp_filtered": 67.37, "supplement_filtered": 57.89, "ewok_filtered": 50.34, "entity_tracking": 19.55, "comps": 51.89, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 48.0 }, "checkpoint": "9a66d00d0021b8689ce44ce9898e1627313349efa4bd4946aabb6bd5fda44470", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s50eb37b351b870", "sd162e9a04349f7", "s5195c57f8b3797", "sfd49ecb03e8e06", "s61509c2e16a7cf", "sb8540f9fdc89e7", "sd702222d1cd47e" ] }, { "age": 100, "six": 47.075, "five": 49.66, "tasks": { "blimp_filtered": 67.75, "supplement_filtered": 58.93, "ewok_filtered": 50.23, "entity_tracking": 19.42, "comps": 51.97, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 45.0 }, "checkpoint": "6616f6c885635155005b304716cdbed7baca87e95d0f58b6adea79c9020e0fbf", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s50eb37b351b870", "sd162e9a04349f7", "s66c671806b00a4", "s9794084a50c2dd", "s61509c2e16a7cf", "sb8540f9fdc89e7", "sd702222d1cd47e" ] } ], "peak": { "five": { "score": 49.66, "ages": [ 100 ], "rank": 168 }, "six": { "score": 47.3575, "ages": [ 90 ], "rank": 155 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 130, "id": "run/IPRLM_MOS3_320_LAST2_SEED731946281_40M_20261013", "original_name": "IPRLM_MOS3_320_LAST2_SEED731946281_40M_20261013", "title": "320幅・FFN1120・9層", "topology": "MoS3・共通2層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last2", "attention_received_gate": true, "output_components": 3, "output_state_layers": 2, "reference_attention_width": 256, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "After40M training: evaluate40M only", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.715, "comparison_scope": "Same parent320 architecture/LR/data order; one alternate seed; common70/80/90/100M peak and individual tasks. Two runs do not estimate seed variance precisely." }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 155, "sources": [ "s763620ffa05e28", "s14a1886a593fc4", "s11c8bf527f8870" ], "points": [ { "age": 40.0, "six": 47.348333333333336, "five": 50.076, "tasks": { "blimp_filtered": 68.07, "supplement_filtered": 58.37, "ewok_filtered": 52.0, "entity_tracking": 19.42, "comps": 52.52, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 48.0 }, "checkpoint": "75cf9b34087b97ca0f37625feba78ce8390e63fcd26878e3544f59645e47e32d", "precision": "fp32", "verification": "primary", "sources": [ "sc74dccc1111e32", "s182673791d7b82" ] } ], "peak": { "five": { "score": 50.076, "ages": [ 40.0 ], "rank": 144 }, "six": { "score": 47.348333333333336, "ages": [ 40.0 ], "rank": 156 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 131, "id": "baby/vistarc_original256_received_gate_20261005", "original_name": "元256・受信ゲート追加", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "あり", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_received_gate": true, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 38, "sources": [ "s2dfae00569b2d9", "s4429972b8a6ac9", "sb3469368c88cfb", "sf562d30129eb86", "s409109b8ba4506", "s9f854ef7f7f322", "s08db7fda5d51ce" ], "points": [ { "age": 80, "six": 47.345, "five": 49.672, "tasks": { "blimp_filtered": 66.46, "supplement_filtered": 58.56, "ewok_filtered": 50.71, "entity_tracking": 20.3, "comps": 52.33, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 52.0 }, "checkpoint": "c0543280987179c3214e88629bbc5e339f2116309be64cc44680f114738665a4", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s63f9c908110207", "s5c1fad92338c27", "s46ac819a4e279c", "s121359d280c968", "s1afc22962b03ec", "s0ebd952533c0f0", "s770b705d4729f7" ] }, { "age": 100, "six": 46.385, "five": 48.826, "tasks": { "blimp_filtered": 67.2, "supplement_filtered": 54.99, "ewok_filtered": 50.86, "entity_tracking": 19.19, "comps": 51.89, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 47.0 }, "checkpoint": "abdf7e993dcdfd2733150fc974c42afe9ffb41d8cc0a0328a28bb0194a9d4c67", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5ecfc5079069e7", "s5c1fad92338c27", "s3797e6d9f4bdbf", "s121359d280c968", "s1afc22962b03ec", "s0ebd952533c0f0", "s770b705d4729f7" ] } ], "peak": { "five": { "score": 49.672, "ages": [ 80 ], "rank": 167 }, "six": { "score": 47.345, "ages": [ 80 ], "rank": 157 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 132, "id": "run/IPRLM352_FFN1232_ADDITIVE_SEED20260907_20261016", "original_name": "IPRLM352_FFN1232_ADDITIVE_SEED20260907_20261016", "title": "352幅・FFN1232・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 352, "ffn": 1232, "depth": 9, "parts": { "width": 352, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 352, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M parent320/additive320/additive336; broader peak record separate" }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 95, "sources": [ "s71b4e5cfaca189", "scc6d67fceeb9f4", "s4bca52ef55e2e3" ], "points": [ { "age": 40.0, "six": 47.3425, "five": 50.166, "tasks": { "blimp_filtered": 68.97, "supplement_filtered": 57.82, "ewok_filtered": 52.25, "entity_tracking": 19.45, "comps": 52.34, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 48.0 }, "checkpoint": "f1aa5c4908af94580561edb629d905e57b3aa71c37814434807ccb64009deb96", "precision": "fp32", "verification": "primary", "sources": [ "s699e761d15e54f", "sace5422dff8b3f" ] } ], "peak": { "five": { "score": 50.166, "ages": [ 40.0 ], "rank": 135 }, "six": { "score": 47.3425, "ages": [ 40.0 ], "rank": 158 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 133, "id": "baby/B_READER_DEPTH10_BATCH4096_BETA2_100M", "original_name": "旧256・10層・単一出力", "title": "256幅・FFN896・10層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 10, "parts": { "width": 256, "ffn": 896, "depth": 10, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "未確認", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 10, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 14, "sources": [ "s0fe436fd0a45c6", "saffc9d1867ab78", "sfc203da675a540", "s9e9c203870a77d", "sa5e62da0ded420", "sb37d5dc6021ef5", "s86280e1cb21884", "sfa1b39e3f2bee7" ], "points": [ { "age": 80, "six": 47.340833333333336, "five": 49.582, "tasks": { "blimp_filtered": 64.77, "supplement_filtered": 57.28, "ewok_filtered": 52.75, "entity_tracking": 21.14, "comps": 51.97, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 48.0 }, "checkpoint": "89af1a4f05452f32ddde15c0b4644c62ea9a7d91503b78dd5e16021eae8a9675", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s131cc22d8ef627", "s5e4b28558c9b61", "s95f651c8b7bcd2", "s9233290945a939", "sd8b7680fe13844", "s3d8f92bb8a7edf", "s142f47a4b7b127", "s2f573b9c456ec2", "s40b607674115e0", "s53233c6ba2edbf", "s51d6cd28bff050" ] }, { "age": 100, "six": 46.45333333333333, "five": 49.402, "tasks": { "blimp_filtered": 65.74, "supplement_filtered": 56.42, "ewok_filtered": 52.43, "entity_tracking": 20.38, "comps": 52.04, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 44.0 }, "checkpoint": "0a5ec0b328a4032b2303d0d86b034dd11933834264051f014c885eb7312b95c2", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s131cc22d8ef627", "s5e4b28558c9b61", "sb7be9899c87f5f", "s564501e2c3b380", "s3d8f92bb8a7edf", "s142f47a4b7b127", "s2f573b9c456ec2", "s40b607674115e0", "s53233c6ba2edbf", "s51d6cd28bff050" ] } ], "peak": { "five": { "score": 49.582, "ages": [ 80 ], "rank": 171 }, "six": { "score": 47.340833333333336, "ages": [ 80 ], "rank": 159 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 134, "id": "run/IPRLM384_MOS_OUTPUT_LR95_SEED20260907_40M_20261020", "original_name": "IPRLM384_MOS_OUTPUT_LR95_SEED20260907_40M_20261020", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "MoS出力8行列LR−5%（FFN入口・出口一致版ベース）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005333333333333334, 0.00046666666666666666, 0.00039999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All11matchedseed40MLRconditions;allgenres/pairedoutcomes/fullinventory" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 135, "sources": [ "s7db8b936bb812b", "s39ef4e4504e083" ], "points": [ { "age": 40.0, "six": 47.335, "five": 49.76, "tasks": { "blimp_filtered": 69.38, "supplement_filtered": 57.78, "ewok_filtered": 48.97, "entity_tracking": 20.55, "comps": 52.12, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 51.0 }, "checkpoint": "f2a593bcf44f726ea4705e15e55a6b15fcf059c2b81335119dee1da7e87e39cc", "precision": "fp32", "verification": "primary", "sources": [ "s0a5d2bf94d541f", "sb77c75a7b800a9" ] } ], "peak": { "five": { "score": 49.76, "ages": [ 40.0 ], "rank": 162 }, "six": { "score": 47.335, "ages": [ 40.0 ], "rank": 160 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 135, "id": "baby/B_REPAIRED8_LONG512_HPARAM_100M", "original_name": "修正版8層・添付512設定", "title": "256幅・FFN896・8層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 8, "parts": { "width": 256, "ffn": 896, "depth": 8, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "未確認", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 8, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "vocab": 16384, "seed": 20260907, "reader_feature_mode": "bounded_innovation", "innovation_variance_floor": 0.1, "spread_direction_floor": 0.01, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "Save each10M without stopping training; score80M and100M after100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact FP32 eager CUDA graphs; TF32 disabled; choose validated score batching and graph chunk", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "checkpoint_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 68905, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "5167571743565769b6fdc7387ad101402745fffe122104c8e0bb04a5af851ad4", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "405af3d2d6ef1e2825c2e26d919bb21785b47615144941fad986194f66900174", "grouping": "Adjacent original lines, target256 content tokens, max512 except individual long lines, max32 lines; reset at source, visible markers and empty lines", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "context_block_content_tokens_mean": 235.85783324867572, "effective_packed_context_tokens_mean_per_epoch": [ 162.6161508924762, 162.64035644821973, 162.64681248015242, 162.62744592180988, 162.65326902465165, 162.6096972944013, 162.6355147606053, 162.64035644821973, 162.62905962551721, 162.6258322501265 ] }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": false, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "gpu_power_limit_watts": 260 }, "condition_conflicts": [], "profile": 28, "sources": [ "s1a68d8025fa9c5", "s0d51c2806bdf9d", "sc94f27d4b00ce1", "s2c76c6cf3591ec", "s31b46470f0f5f4", "s9b86a4e1bac4d6", "s1978aef90a60f3" ], "points": [ { "age": 80, "six": 46.465833333333336, "five": 49.714, "tasks": { "blimp_filtered": 68.33, "supplement_filtered": 57.29, "ewok_filtered": 51.11, "entity_tracking": 19.92, "comps": 51.92, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 42.0 }, "checkpoint": "6a506a9b2fc5c1e20282a0761e86ff05d984a202e58e201a756160d6cd3d50a9", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s2fd508ff1e606a", "sf492eee741bd74", "s80cda9de16d2eb", "s43aad0b29ebc37", "sd7b4a05437428d", "s738d9cc1512201", "s5eb57869f6ec82", "s73e0bc9f73041f", "s43bdf806380d42" ] }, { "age": 100, "six": 47.3025, "five": 50.124, "tasks": { "blimp_filtered": 68.39, "supplement_filtered": 59.06, "ewok_filtered": 51.48, "entity_tracking": 19.73, "comps": 51.96, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "076c93665700b0c3ae2b65c42bab8f26d1b261f6a988c83ffd910ccfd5c8b826", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s2fd508ff1e606a", "sf492eee741bd74", "sb922b3d80e54c3", "s65015d08a1a529", "sd7b4a05437428d", "s738d9cc1512201", "s5eb57869f6ec82", "s73e0bc9f73041f", "s43bdf806380d42" ] } ], "peak": { "five": { "score": 50.124, "ages": [ 100 ], "rank": 142 }, "six": { "score": 47.3025, "ages": [ 100 ], "rank": 163 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 136, "id": "run/VISTA336_CANDIDATE_READ_RECEIVED_GATE_FFN_ADD_SEED20260907_20261017", "original_name": "VISTA336_CANDIDATE_READ_RECEIVED_GATE_FFN_ADD_SEED20260907_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": "未確認", "received": "あり", "output": "単一出力", "reference": "9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTA_CANDIDATE_READ_RECEIVED_GATE_FFN_ADD", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": true, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": true, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "40M only after firstseed training; no remainingseed auto-launch", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 with bounded live graph/KV cache; validated native gate included", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus current candidate-read baseline; only Attention received gate added." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 160, "sources": [ "sea0e3731acaae3", "sa4636c47e7a304", "sceb9c9c75534e0" ], "points": [ { "age": 40.0, "six": 47.285, "five": 49.418, "tasks": { "blimp_filtered": 68.71, "supplement_filtered": 56.6, "ewok_filtered": 51.34, "entity_tracking": 18.69, "comps": 51.75, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 48.0 }, "checkpoint": "854a90c3d731541e25d701906d42c68f79a818b9b7ef9f6a75071058c534b8fe", "precision": "fp32", "verification": "primary", "sources": [ "s788cdd9f5e3822", "s44aa68ecc36410" ] } ], "peak": { "five": { "score": 49.418, "ages": [ 40.0 ], "rank": 176 }, "six": { "score": 47.285, "ages": [ 40.0 ], "rank": 164 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 137, "id": "baby/B_CORRECTED_100M", "original_name": "修正B", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "local six-axis benchmark at every 10M checkpoint; no Reading or official submission evaluation", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 1104106, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "611332e27d15ac01cd3f65cce4c09b5c291fd0150732ed2c134cc21abf4ba121", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_ORDER_CONTENT_NO_OMISSIONS", "boundary_policy": "document-isolated causal masks; shuffled each pass; all prefixes and tails retained", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 3, "sources": [ "sda1c846064c240", "sac704a5834af4c", "s69737d39777c42", "sfb5c4632a707b4", "sc91a5ad427c07a", "s4346b5606cf68a", "sd6f570fb30a005" ], "points": [ { "age": 10, "six": 45.700833333333335, "five": 47.908, "tasks": { "blimp_filtered": 66.2, "supplement_filtered": 56.91, "ewok_filtered": 49.29, "entity_tracking": 17.94, "comps": 49.2, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "abfe87074e266853b02c45a070788cfa3371a63235b693b5b9fef8a4b7291b6d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1b884cfc191c01", "sa52d0a3385f6c7", "sedb09ee39c3e8a", "sc96f06f9102133", "sd2ce7fbe615302", "s3009f759e5eba1", "s94eb0557d18dfe" ] }, { "age": 20, "six": 46.07083333333333, "five": 48.446, "tasks": { "blimp_filtered": 67.01, "supplement_filtered": 55.77, "ewok_filtered": 49.93, "entity_tracking": 19.02, "comps": 50.5, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 48.0 }, "checkpoint": "995f697c5953a99eba4fac58a8ea29a9aef88b46393d1c5ed33b4657712c608d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1b884cfc191c01", "sa52d0a3385f6c7", "s01b94a9f2be585", "sc482a2b16856d7", "sd2ce7fbe615302", "s3009f759e5eba1", "s94eb0557d18dfe" ] }, { "age": 30, "six": 45.90083333333333, "five": 48.254, "tasks": { "blimp_filtered": 66.9, "supplement_filtered": 55.13, "ewok_filtered": 50.15, "entity_tracking": 17.65, "comps": 51.44, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 44.0 }, "checkpoint": "5723044be99ccae8fe9ea94820ca9bfa0992d94aafd1b9c3ab4bb550d06a0148", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1b884cfc191c01", "sa52d0a3385f6c7", "s4e108a5c1f8294", "s289c71c65b2c3c", "sd2ce7fbe615302", "s3009f759e5eba1", "s94eb0557d18dfe" ] }, { "age": 40, "six": 46.74666666666667, "five": 48.566, "tasks": { "blimp_filtered": 67.9, "supplement_filtered": 54.96, "ewok_filtered": 50.71, "entity_tracking": 18.11, "comps": 51.15, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 52.0 }, "checkpoint": "5ba930fcf7daca00780651dec4eab09daef143437f860063d72b66a5b9435187", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1b884cfc191c01", "sa52d0a3385f6c7", "s38b0377a8d782b", "sccd27f1fdd9e6c", "sd2ce7fbe615302", "s3009f759e5eba1", "s94eb0557d18dfe" ] }, { "age": 50, "six": 46.25083333333333, "five": 48.568, "tasks": { "blimp_filtered": 68.32, "supplement_filtered": 56.37, "ewok_filtered": 49.66, "entity_tracking": 17.38, "comps": 51.11, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "4a758159f8ec60010f4cd86b30e8702dece00d157620612fc546f8fa60da09d9", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1b884cfc191c01", "sa52d0a3385f6c7", "s74df3066e334a3", "sb9412ba671f250", "sd2ce7fbe615302", "s3009f759e5eba1", "s94eb0557d18dfe" ] }, { "age": 60, "six": 46.5625, "five": 49.342, "tasks": { "blimp_filtered": 68.51, "supplement_filtered": 57.22, "ewok_filtered": 50.9, "entity_tracking": 18.45, "comps": 51.63, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 43.0 }, "checkpoint": "82fd38fe46bb2985b80ee5ec95ee288dc0716ccc05edb47caba46a041c1525d0", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1b884cfc191c01", "sa52d0a3385f6c7", "sa7205d36529b6c", "sfab78c49cf531a", "sd2ce7fbe615302", "s3009f759e5eba1", "s94eb0557d18dfe" ] }, { "age": 70, "six": 46.5425, "five": 49.106, "tasks": { "blimp_filtered": 67.76, "supplement_filtered": 56.88, "ewok_filtered": 50.87, "entity_tracking": 18.59, "comps": 51.43, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 49.0 }, "checkpoint": "9c74b79b5ad54a15232b1962eb2db6efbb87aaec0e172848e7a1a3c2ea3b6b73", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1b884cfc191c01", "sa52d0a3385f6c7", "s2647c693d25ddc", "s73e6c81609294c", "sd2ce7fbe615302", "s3009f759e5eba1", "s94eb0557d18dfe" ] }, { "age": 80, "six": 47.251666666666665, "five": 49.266, "tasks": { "blimp_filtered": 67.79, "supplement_filtered": 57.49, "ewok_filtered": 50.86, "entity_tracking": 18.63, "comps": 51.56, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 53.0 }, "checkpoint": "4b22933b951ab341bb5f4664ef0ebe412a1cca3c265247cfd45abf040b8d88c2", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1b884cfc191c01", "sa52d0a3385f6c7", "s3ad91f2976064d", "s8ad134fa2de4fa", "sd2ce7fbe615302", "s3009f759e5eba1", "s94eb0557d18dfe" ] }, { "age": 90, "six": 46.468333333333334, "five": 48.82, "tasks": { "blimp_filtered": 67.43, "supplement_filtered": 58.24, "ewok_filtered": 49.87, "entity_tracking": 17.17, "comps": 51.39, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 50.0 }, "checkpoint": "bc55a68871c57e20aec15cd08e914ca7cb23bc0b57cd7ef657c099673c78835d", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1b884cfc191c01", "sa52d0a3385f6c7", "s241beefe33ff7f", "sa3c6aba440ec3c", "sd2ce7fbe615302", "s3009f759e5eba1", "s94eb0557d18dfe" ] }, { "age": 100, "six": 46.3825, "five": 48.826, "tasks": { "blimp_filtered": 67.68, "supplement_filtered": 57.35, "ewok_filtered": 49.47, "entity_tracking": 17.99, "comps": 51.64, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "b7087b8182daf49b31a92d3f6d1b9e2f17f5e9c555177728c7158c31ed92c549", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1b884cfc191c01", "sa52d0a3385f6c7", "s35109033df8af5", "s2923a1f816cdce", "sd2ce7fbe615302", "s3009f759e5eba1", "s94eb0557d18dfe" ] } ], "peak": { "five": { "score": 49.342, "ages": [ 60 ], "rank": 178 }, "six": { "score": 47.251666666666665, "ages": [ 80 ], "rank": 165 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 138, "id": "run/IPRLM336_FFN1288_ADDITIVE_SEED731946281_20261017", "original_name": "IPRLM336_FFN1288_ADDITIVE_SEED731946281_20261017", "title": "336幅・FFN1288・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 336, "ffn": 1288, "depth": 9, "parts": { "width": 336, "ffn": 1288, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.8333333333333335 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 336, "layers": 9, "hidden": 1288, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M parent320, additive320,336FFN1120/1176/1232/1288; later peaks separate." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 92, "sources": [ "s0a0c9d9072d76b", "s270bd9abd72ac1", "s00ac2526fe4aa3" ], "points": [ { "age": 40.0, "six": 47.25, "five": 50.352, "tasks": { "blimp_filtered": 68.89, "supplement_filtered": 58.0, "ewok_filtered": 53.13, "entity_tracking": 19.3, "comps": 52.44, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 46.0 }, "checkpoint": "bff2b5af401aaeae315ab640fcfb0f9e676695a31d4056c1ebdca15748e92930", "precision": "fp32", "verification": "primary", "sources": [ "sa1c6bbe7eb30d8", "sb1eef2caf62642" ] } ], "peak": { "five": { "score": 50.352, "ages": [ 40.0 ], "rank": 124 }, "six": { "score": 47.25, "ages": [ 40.0 ], "rank": 166 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 139, "id": "run/IPRLM336_FFN1288_ADDITIVE_SEED20260907_20261017", "original_name": "IPRLM336_FFN1288_ADDITIVE_SEED20260907_20261017", "title": "336幅・FFN1288・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1288, "depth": 9, "parts": { "width": 336, "ffn": 1288, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.8333333333333335 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 336, "layers": 9, "hidden": 1288, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M parent320, additive320,336FFN1120/1176/1232/1288; later peaks separate." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 92, "sources": [ "s64b2d2e1439c03", "scbd5bd32a6b6f8", "sd7c0cdcaad4e5d" ], "points": [ { "age": 40.0, "six": 47.2375, "five": 50.146, "tasks": { "blimp_filtered": 70.16, "supplement_filtered": 58.42, "ewok_filtered": 50.4, "entity_tracking": 19.79, "comps": 51.96, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 45.0 }, "checkpoint": "25d3d50474208051e0460c234f5ae2c3a31a699f9af80a48980546c89ce26f75", "precision": "fp32", "verification": "primary", "sources": [ "s144020edd1be08", "s3ef88ad552162b" ] } ], "peak": { "five": { "score": 50.146, "ages": [ 40.0 ], "rank": 138 }, "six": { "score": 47.2375, "ages": [ 40.0 ], "rank": 167 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 140, "id": "baby/B_CONTEXTUAL_READER_80M_100M", "original_name": "文脈Reader・80Mから継続", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "contextual_reader": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "Save90/100M and score only after100M; local six axes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 1, "sources": [ "s4bcfa0ff6aa39e", "scca13fcdddc303", "s8e8b2dbec73a04", "s4d8d79f2788bf3" ], "points": [ { "age": 90, "six": 47.236666666666665, "five": 49.36, "tasks": { "blimp_filtered": 67.15, "supplement_filtered": 57.36, "ewok_filtered": 50.67, "entity_tracking": 19.79, "comps": 51.83, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 48.0 }, "checkpoint": "f52d8a29d70d0b8b1a0fde8275c2bb4dce59b86043a6c1d511b6afe188fa2be7", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc02c78a8b1bec6", "sb9c91a3163c618", "s016a42a70cccff", "s7e1207f8118154", "s3019219adb6d27", "s39ccbbe0021faf", "sa25788e0e48bc7" ] }, { "age": 100, "six": 46.9825, "five": 49.646, "tasks": { "blimp_filtered": 68.13, "supplement_filtered": 58.59, "ewok_filtered": 50.48, "entity_tracking": 19.1, "comps": 51.93, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 45.0 }, "checkpoint": "94b7d33dea72063ca24cc51b9589b4c8f82b6d03a5233e4b222898a7f027ba8b", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc02c78a8b1bec6", "sb9c91a3163c618", "s560c7c305e7238", "sb6c85bfd0d1a6d", "s3019219adb6d27", "s39ccbbe0021faf", "sa25788e0e48bc7" ] } ], "peak": { "five": { "score": 49.646, "ages": [ 100 ], "rank": 169 }, "six": { "score": 47.236666666666665, "ages": [ 90 ], "rank": 168 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 141, "id": "baby/B_ONE_CYCLE_100M", "original_name": "修正B・系列長1サイクル", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "local six-axis benchmark at every 10M checkpoint; no Reading or official submission evaluation", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 1104106, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "20d5698de003d537524521ed8eabfd03ccd4c779e5641fe6803249707ba8e458", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_ORDER_CONTENT_NO_OMISSIONS", "boundary_policy": "document-isolated causal masks; shuffled each pass; all prefixes and tails retained", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 7, "sources": [ "s109bb6575adcec", "safda63e0d839a1", "s82c01b6ba9052e", "sb117ff174180c0" ], "points": [ { "age": 10, "six": 45.700833333333335, "five": 47.908, "tasks": { "blimp_filtered": 66.2, "supplement_filtered": 56.91, "ewok_filtered": 49.29, "entity_tracking": 17.94, "comps": 49.2, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "b57dff52622a2ecde14503537c5906d1a9dce9f0c161359b4f4c89c1f9c17177", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5e4f6f7d6559c8", "se40d927a1d01aa", "sc067f5ea34ea89", "s9531d5d2eb38d0", "sd4c2053e56d657", "sfa0998c68bed37", "sba0de4ac5dbb54" ] }, { "age": 20, "six": 46.056666666666665, "five": 48.432, "tasks": { "blimp_filtered": 67.49, "supplement_filtered": 56.18, "ewok_filtered": 49.25, "entity_tracking": 18.38, "comps": 50.86, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 47.0 }, "checkpoint": "1e594eed73950ec2a319b62f070bdd79e7ab9efa54452b76d113ac84440e50a4", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5e4f6f7d6559c8", "se40d927a1d01aa", "se9aa30f65ceab7", "s47dd57ce5fdb10", "sd4c2053e56d657", "sfa0998c68bed37", "sba0de4ac5dbb54" ] }, { "age": 30, "six": 46.02333333333333, "five": 48.298, "tasks": { "blimp_filtered": 66.99, "supplement_filtered": 56.11, "ewok_filtered": 49.41, "entity_tracking": 17.92, "comps": 51.06, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 46.0 }, "checkpoint": "cecb1e4851f83fc1afd2f31859b7d624e548aa70a908f26bac9eb203075bddb0", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5e4f6f7d6559c8", "se40d927a1d01aa", "sd3fcaa25410d1e", "s9d1a5334635fb9", "sd4c2053e56d657", "sfa0998c68bed37", "sba0de4ac5dbb54" ] }, { "age": 40, "six": 46.395, "five": 48.65, "tasks": { "blimp_filtered": 68.19, "supplement_filtered": 55.23, "ewok_filtered": 50.21, "entity_tracking": 18.21, "comps": 51.41, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 45.0 }, "checkpoint": "78dadc747a775b55d0a4ca19f11d7d3fa112a96ba000111c2af4ed3a3c182935", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5e4f6f7d6559c8", "se40d927a1d01aa", "s1cce94815cbc48", "s63378f4171496f", "sd4c2053e56d657", "sfa0998c68bed37", "sba0de4ac5dbb54" ] }, { "age": 50, "six": 46.6725, "five": 49.28, "tasks": { "blimp_filtered": 69.15, "supplement_filtered": 56.34, "ewok_filtered": 51.0, "entity_tracking": 18.43, "comps": 51.48, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 43.0 }, "checkpoint": "b49d9e10213a58a90aacab76139be868753e3d7d0a769af14201c623cffc28c8", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5e4f6f7d6559c8", "se40d927a1d01aa", "s40b02ea4bba25c", "s290e76d1a91a9d", "sd4c2053e56d657", "sfa0998c68bed37", "sba0de4ac5dbb54" ] }, { "age": 60, "six": 46.3775, "five": 48.72, "tasks": { "blimp_filtered": 68.77, "supplement_filtered": 53.84, "ewok_filtered": 50.75, "entity_tracking": 18.63, "comps": 51.61, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "973cc4074bde9cc8eb6c72d7d73717cfae8e68b949a23acef0d989601e9442ba", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5e4f6f7d6559c8", "se40d927a1d01aa", "sc46f87171b02ad", "sa15d3546ea0185", "sd4c2053e56d657", "sfa0998c68bed37", "sba0de4ac5dbb54" ] }, { "age": 70, "six": 47.19583333333333, "five": 48.908, "tasks": { "blimp_filtered": 68.62, "supplement_filtered": 54.92, "ewok_filtered": 50.45, "entity_tracking": 18.99, "comps": 51.56, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 53.0 }, "checkpoint": "a66670bbb9a671e04055e04776290a021d7ce0040ddaa8409037bfd25fb787ff", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5e4f6f7d6559c8", "se40d927a1d01aa", "s7514e34084c478", "s0185ebb513e6c5", "sd4c2053e56d657", "sfa0998c68bed37", "sba0de4ac5dbb54" ] }, { "age": 80, "six": 47.21333333333333, "five": 49.726, "tasks": { "blimp_filtered": 68.96, "supplement_filtered": 58.27, "ewok_filtered": 51.09, "entity_tracking": 19.07, "comps": 51.24, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 46.0 }, "checkpoint": "dee375b7c1736ca77be0693ce211da3a4b1bcf798d0b4e1996a555eb5ee3d2c9", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5e4f6f7d6559c8", "se40d927a1d01aa", "sbee91e4a973e35", "s2a7e026d8b2ed9", "sd4c2053e56d657", "sfa0998c68bed37", "sba0de4ac5dbb54" ] }, { "age": 90, "six": 46.43333333333333, "five": 48.99, "tasks": { "blimp_filtered": 68.78, "supplement_filtered": 56.42, "ewok_filtered": 49.8, "entity_tracking": 18.6, "comps": 51.35, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 44.0 }, "checkpoint": "7155d262c798fa57a53fd400272d5e7a8d930e45a4d311a76286ca3cfaaae953", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5e4f6f7d6559c8", "se40d927a1d01aa", "sd7f6a7fb8c8af5", "s27124328a9a931", "sd4c2053e56d657", "sfa0998c68bed37", "sba0de4ac5dbb54" ] }, { "age": 100, "six": 46.71083333333333, "five": 48.926, "tasks": { "blimp_filtered": 68.49, "supplement_filtered": 55.99, "ewok_filtered": 49.76, "entity_tracking": 19.04, "comps": 51.35, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 47.0 }, "checkpoint": "826a318c81ea2ce7575c1a529e650fbd32207d766162797e48aeb01f97f5752b", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s5e4f6f7d6559c8", "se40d927a1d01aa", "sd987940fdef1b0", "se0b32856f7bcf3", "sd4c2053e56d657", "sfa0998c68bed37", "sba0de4ac5dbb54" ] } ], "peak": { "five": { "score": 49.726, "ages": [ 80 ], "rank": 164 }, "six": { "score": 47.21333333333333, "ages": [ 80 ], "rank": 169 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 142, "id": "baby/B_REPAIRED8_BABY_HPARAM_100M", "original_name": "修正版8層・従来Baby設定", "title": "256幅・FFN896・8層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 8, "parts": { "width": 256, "ffn": 896, "depth": 8, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "未確認", "distance": "未確認", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 8, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "vocab": 16384, "seed": 20260907, "reader_feature_mode": "bounded_innovation", "innovation_variance_floor": 0.1, "spread_direction_floor": 0.01, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "Save each10M without stopping training; score80M and100M after100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact FP32 eager CUDA graphs; TF32 disabled; choose validated score batching and graph chunk", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "checkpoint_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": false, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "gpu_power_limit_watts": 260 }, "condition_conflicts": [], "profile": 27, "sources": [ "s159106b7dcf5e8", "sb14cf7ef6b19dc", "sca22fd032ed23a", "saaf91802439905", "s45d29f430542c1", "s9a0ebbd0fef665", "s6f7473fba4cfa0" ], "points": [ { "age": 80, "six": 47.21333333333333, "five": 49.82, "tasks": { "blimp_filtered": 66.34, "supplement_filtered": 56.34, "ewok_filtered": 53.99, "entity_tracking": 20.44, "comps": 51.99, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 47.0 }, "checkpoint": "1b297dd15fdb397211e5cbf92f5f4e69d49d4531e5b43deab70ad4602e2e9358", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4454735eb94a8d", "s8b7106a2dba428", "s8356f0f285d62f", "s9e603641bbdee1", "scb18183f556242", "s1ef393fc61bef6", "sf06e4ecf6b3605", "s9f0c5e60085b3f", "sce3173f9b53b49" ] }, { "age": 100, "six": 46.79083333333333, "five": 49.51, "tasks": { "blimp_filtered": 65.17, "supplement_filtered": 57.82, "ewok_filtered": 51.84, "entity_tracking": 20.6, "comps": 52.12, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "f0d3ca5b232ac28205affb861c7d9a9e78a67cd9da31ac396ebae99c48db72fe", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4454735eb94a8d", "s8b7106a2dba428", "s785eeeca595944", "s942ac60cef59cd", "scb18183f556242", "s1ef393fc61bef6", "sf06e4ecf6b3605", "s9f0c5e60085b3f", "sce3173f9b53b49" ] } ], "peak": { "five": { "score": 49.82, "ages": [ 80 ], "rank": 159 }, "six": { "score": 47.21333333333333, "ages": [ 80 ], "rank": 169 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 143, "id": "baby/O_READER48_ORIGINAL_ATTN_Q0", "original_name": "Reader48・元Attention・Q0更新", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "なし", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": false, "qk_norm_eps": 1e-06, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 29, "sources": [ "sba48a852208bed", "sf502b016586db3", "sdb62cc469c0e59", "s74071cf46ce0a0", "s4a3f7a1d9d7d8a", "sac498c88d337e0", "s23fcdef94c3a06" ], "points": [ { "age": 80, "six": 47.204166666666666, "five": 49.706, "tasks": { "blimp_filtered": 65.9, "supplement_filtered": 59.02, "ewok_filtered": 51.16, "entity_tracking": 19.99, "comps": 52.46, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 49.0 }, "checkpoint": "4e30ddcb0f956689ac41a388ee814ee34123b062261c964c8a0e97617c8789b7", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s35d6a55238ce09", "s7b05493e1e601f", "s9f13ec3a8227cc", "sf7154e06199641", "sac3eb35b593a77", "se87e5bb56dd23e" ] }, { "age": 100, "six": 46.961666666666666, "five": 49.23, "tasks": { "blimp_filtered": 66.75, "supplement_filtered": 56.71, "ewok_filtered": 50.82, "entity_tracking": 20.33, "comps": 51.54, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 46.0 }, "checkpoint": "3b9072e0759a5f39810ab1bf9b43de00bc2c5dc3287168dced3e8932811083a1", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s35d6a55238ce09", "s7b05493e1e601f", "s468f11ddf1b20b", "sff33da5afa3400", "sac3eb35b593a77", "se87e5bb56dd23e" ] } ], "peak": { "five": { "score": 49.706, "ages": [ 80 ], "rank": 165 }, "six": { "score": 47.204166666666666, "ages": [ 80 ], "rank": 171 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 144, "id": "run/IPRLM384_FFN1408_LRPLUS10_ALL110_SEED20260907_40M_20261020", "original_name": "IPRLM384_FFN1408_LRPLUS10_ALL110_SEED20260907_40M_20261020", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "全LRを1割増", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005866666666666668, 0.0005133333333333333, 0.00043999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "aftereachcase40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Three LR allocation cases and frozen baseline40M; full inventory; no final ranking from40M alone" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 115, "sources": [ "sf81f1d4e10620c", "s1b2fda962934e4" ], "points": [ { "age": 40.0, "six": 47.2025, "five": 49.504, "tasks": { "blimp_filtered": 66.27, "supplement_filtered": 57.07, "ewok_filtered": 51.64, "entity_tracking": 20.48, "comps": 52.06, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 51.0 }, "checkpoint": "a06006e7d009d65b92b56a46fb6ec57fb682a0c0f018cd0b3a406c1a1462b9fc", "precision": "fp32", "verification": "primary", "sources": [ "saf55e692786218", "s79978ca3036470" ] } ], "peak": { "five": { "score": 49.504, "ages": [ 40.0 ], "rank": 174 }, "six": { "score": 47.2025, "ages": [ 40.0 ], "rank": 172 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 145, "id": "baby/B_LR_COSINE_60M_100M", "original_name": "修正B・60M以降cosine", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "local six-axis benchmark at every 10M checkpoint; no Reading or official submission evaluation", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 1104106, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "611332e27d15ac01cd3f65cce4c09b5c291fd0150732ed2c134cc21abf4ba121", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_ORDER_CONTENT_NO_OMISSIONS", "boundary_policy": "document-isolated causal masks; shuffled each pass; all prefixes and tails retained", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 6, "sources": [ "se396bf843498f2", "s3b135271634c7b", "s7147f9b8140ad0", "s159331e205e9fd", "s4519f6fd9906fb", "s75e810455461e0", "s83ff83a7a71faf" ], "points": [ { "age": 10, "six": 45.700833333333335, "five": 47.908, "tasks": { "blimp_filtered": 66.2, "supplement_filtered": 56.91, "ewok_filtered": 49.29, "entity_tracking": 17.94, "comps": 49.2, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "80d3faea15ba3f399dcaf42d6003fe8acaf6008c1a22efdd88c8bc645cf5a869", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc7416fa3abc9f6", "sd67246c0ee71d0", "s3da4c5a4b53fef", "s7bf54e0ec5ff42", "sa268e5f8bc4a6f", "s134032805f42a3", "s8e56e27539302d" ] }, { "age": 20, "six": 46.07083333333333, "five": 48.446, "tasks": { "blimp_filtered": 67.01, "supplement_filtered": 55.77, "ewok_filtered": 49.93, "entity_tracking": 19.02, "comps": 50.5, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 48.0 }, "checkpoint": "61876ff52c4fea6c540bd908851079c36e7f8e49e90847abb91733e3d1f0b448", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc7416fa3abc9f6", "sd67246c0ee71d0", "s8a12ba14fbfa31", "se5df4d494bc96f", "sa268e5f8bc4a6f", "s134032805f42a3", "s8e56e27539302d" ] }, { "age": 30, "six": 45.90083333333333, "five": 48.254, "tasks": { "blimp_filtered": 66.9, "supplement_filtered": 55.13, "ewok_filtered": 50.15, "entity_tracking": 17.65, "comps": 51.44, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 44.0 }, "checkpoint": "47559fac2a6e9b400e19407321811b06caf24cd4b5ab76905974a445890e0702", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc7416fa3abc9f6", "sd67246c0ee71d0", "sed84f55d36d8a6", "s07783c0b2440ec", "sa268e5f8bc4a6f", "s134032805f42a3", "s8e56e27539302d" ] }, { "age": 40, "six": 46.74666666666667, "five": 48.566, "tasks": { "blimp_filtered": 67.9, "supplement_filtered": 54.96, "ewok_filtered": 50.71, "entity_tracking": 18.11, "comps": 51.15, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 52.0 }, "checkpoint": "1f94f1fe2f4bcf43feb802e4152ec53d13b25c2a87a3843ca2a5c2baa8793923", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc7416fa3abc9f6", "sd67246c0ee71d0", "s33b8e9e671befb", "s07c69f55b9fc0b", "sa268e5f8bc4a6f", "s134032805f42a3", "s8e56e27539302d" ] }, { "age": 50, "six": 46.25083333333333, "five": 48.568, "tasks": { "blimp_filtered": 68.32, "supplement_filtered": 56.37, "ewok_filtered": 49.66, "entity_tracking": 17.38, "comps": 51.11, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "a7f497c02d02bc376a4174ae8175abf5968eb52885f38a42f28c536983802a90", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc7416fa3abc9f6", "sd67246c0ee71d0", "s4029c705eef1b2", "s78685b9ad8b044", "sa268e5f8bc4a6f", "s134032805f42a3", "s8e56e27539302d" ] }, { "age": 60, "six": 46.5625, "five": 49.342, "tasks": { "blimp_filtered": 68.51, "supplement_filtered": 57.22, "ewok_filtered": 50.9, "entity_tracking": 18.45, "comps": 51.63, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 43.0 }, "checkpoint": "59b263738e2f681d8518a9f11ba8d873010ae2372eb663f5ce0592547f718f0c", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc7416fa3abc9f6", "sd67246c0ee71d0", "s2c2a5894071e68", "s2749c127881bb2", "sa268e5f8bc4a6f", "s134032805f42a3", "s8e56e27539302d" ] }, { "age": 70, "six": 46.73166666666667, "five": 49.236, "tasks": { "blimp_filtered": 67.93, "supplement_filtered": 57.17, "ewok_filtered": 51.02, "entity_tracking": 18.75, "comps": 51.31, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 49.0 }, "checkpoint": "c84912284218981e5aa44be721786f405d75663553094733c2b8b49fc747618b", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc7416fa3abc9f6", "sd67246c0ee71d0", "sb346620748d4c6", "sa8058b2a98d2a6", "sa268e5f8bc4a6f", "s134032805f42a3", "s8e56e27539302d" ] }, { "age": 80, "six": 47.1575, "five": 49.456, "tasks": { "blimp_filtered": 68.59, "supplement_filtered": 57.94, "ewok_filtered": 50.68, "entity_tracking": 18.53, "comps": 51.54, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 49.0 }, "checkpoint": "5b4e3c06f50c884741f41156f92f535f1ab38511849a5691e24f67771384468b", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc7416fa3abc9f6", "sd67246c0ee71d0", "s276d9f5bcf637d", "sd67535dd3d7800", "sa268e5f8bc4a6f", "s134032805f42a3", "s8e56e27539302d" ] }, { "age": 90, "six": 47.01583333333333, "five": 49.58, "tasks": { "blimp_filtered": 68.53, "supplement_filtered": 58.42, "ewok_filtered": 50.78, "entity_tracking": 18.7, "comps": 51.47, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 48.0 }, "checkpoint": "3729ec5f6f2098b1842159fa99b80b51125d9b0174e405b79ad98221ef769f7c", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc7416fa3abc9f6", "sd67246c0ee71d0", "s86f446f9f0a56e", "s77013e18c21c0c", "sa268e5f8bc4a6f", "s134032805f42a3", "s8e56e27539302d" ] }, { "age": 100, "six": 46.5425, "five": 49.212, "tasks": { "blimp_filtered": 68.45, "supplement_filtered": 57.27, "ewok_filtered": 50.41, "entity_tracking": 18.44, "comps": 51.49, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "ec29b61183ff4e5d80c37a043650690a327081f301426d97ccbc90766a11b6fd", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc7416fa3abc9f6", "sd67246c0ee71d0", "se044ea217254d5", "s5ebbcfad416595", "sa268e5f8bc4a6f", "s134032805f42a3", "s8e56e27539302d" ] } ], "peak": { "five": { "score": 49.58, "ages": [ 90 ], "rank": 172 }, "six": { "score": 47.1575, "ages": [ 80 ], "rank": 173 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 146, "id": "baby/B_CONTEXT_AUX_CYCLE_100M", "original_name": "文脈＋補助loss・巡回", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み・保存仕様の差あり", "architecture": { "id": "VISTARC", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "first_pass_loss_weight": 0.5 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "local six-axis benchmark at every 10M checkpoint; no Reading or official submission evaluation", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "retain_backward_graph": true, "compile_output_tiles": false, "retained_backward_policy": "Eager Transformer body and output tiles: full packed batch retained backward corrupts compiled body gradients. Preserve vectorized fused auxiliary head and CUDA fused AdamW.", "compile_hidden": false, "output_head_memory_policy": "8192-position tiles; recompute feedback head only; retain first/final CE intermediates", "body_memory_policy": "Non-reentrant checkpoint of block0 only at sequence length512; preserve dropout RNG; no block checkpoint at128/256", "speed_revision": "Remove redundant CE and Transformer recomputation; CUDA loss and all parameter gradients bit-exact at dropout0.05 and task weights40:60/60:40" }, "condition_conflicts": [ "source:sf1e73f1536c8cd", "source:sa55acd830a2259" ], "profile": 2, "sources": [ "sfec2f9ddbede4d", "sde56b26b92f886", "sf1e73f1536c8cd", "scc401f8c390eeb", "sdd87d6c6832699", "sc52b64d6de584b", "sa55acd830a2259", "s9e42b3d7b813de", "sc700ed87e246e7", "s92fc03ed296bbd", "s5847ad83237931" ], "points": [ { "age": 10, "six": 45.8375, "five": 48.166, "tasks": { "blimp_filtered": 64.22, "supplement_filtered": 56.43, "ewok_filtered": 51.23, "entity_tracking": 18.25, "comps": 50.7, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 48.0 }, "checkpoint": "d4cd5e22e56376a34f503626e4fe86402e0f3ef437e0f22a224f3db2398ec6f1", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4e74a56a1c62e3", "s193a70d415430c", "s93cb50da072585", "s6f3438046559a5", "sc2859d1d29b588", "sd6e90f5e06563d", "s78652343516e3f" ] }, { "age": 20, "six": 46.33583333333333, "five": 49.064, "tasks": { "blimp_filtered": 66.13, "supplement_filtered": 57.84, "ewok_filtered": 51.38, "entity_tracking": 18.56, "comps": 51.41, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 45.0 }, "checkpoint": "cc3d6936eef82de2ea2c82dcc59569133a1f97e567be521675c3e72ee2be91cd", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4e74a56a1c62e3", "s193a70d415430c", "s86b812f814561c", "sa9f2042d5ec6b6", "sc2859d1d29b588", "sd6e90f5e06563d", "s78652343516e3f" ] }, { "age": 30, "six": 46.6975, "five": 48.91, "tasks": { "blimp_filtered": 66.6, "supplement_filtered": 57.15, "ewok_filtered": 50.71, "entity_tracking": 18.39, "comps": 51.7, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 47.0 }, "checkpoint": "d13e054169abfda3553f708b48848c03ef830ae2c57121aa0747b1b905d5211b", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4e74a56a1c62e3", "s193a70d415430c", "s2c1feefec651e2", "s5cf044f591eb5e", "sc2859d1d29b588", "sd6e90f5e06563d", "s78652343516e3f" ] }, { "age": 40, "six": 46.44833333333333, "five": 48.914, "tasks": { "blimp_filtered": 67.04, "supplement_filtered": 55.28, "ewok_filtered": 51.56, "entity_tracking": 18.91, "comps": 51.78, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 43.0 }, "checkpoint": "c40abef6bc57cd6ff0d741917a196dde696f5f71e32e14f9e08e0effdcb8e513", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4e74a56a1c62e3", "s193a70d415430c", "s36ec4e64f9d959", "s8f8c00fe5a7997", "sc2859d1d29b588", "sd6e90f5e06563d", "s78652343516e3f" ] }, { "age": 50, "six": 46.15083333333333, "five": 49.036, "tasks": { "blimp_filtered": 66.32, "supplement_filtered": 55.72, "ewok_filtered": 51.94, "entity_tracking": 19.53, "comps": 51.67, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 45.0 }, "checkpoint": "62edea6140bfd901bf3e4f225fff458867397ffa50ccc7a9f0870fd61c10485a", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4e74a56a1c62e3", "s193a70d415430c", "s7cbef233b24226", "s4f9bb2c6f4b46c", "sc2859d1d29b588", "sd6e90f5e06563d", "s78652343516e3f" ] }, { "age": 60, "six": 46.38166666666667, "five": 48.634, "tasks": { "blimp_filtered": 66.99, "supplement_filtered": 54.22, "ewok_filtered": 51.01, "entity_tracking": 19.45, "comps": 51.5, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 45.0 }, "checkpoint": "b0189af69b48e7b068486ece05a2d66cb2555cc0ecad403738213d0dfadcb519", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4e74a56a1c62e3", "s193a70d415430c", "s5294acc2d4bdfb", "sd09d58869bafe1", "sc2859d1d29b588", "sd6e90f5e06563d", "s78652343516e3f" ] }, { "age": 70, "six": 46.535, "five": 48.618, "tasks": { "blimp_filtered": 67.37, "supplement_filtered": 54.55, "ewok_filtered": 50.27, "entity_tracking": 19.44, "comps": 51.46, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 47.0 }, "checkpoint": "095b2bcd271caa77b78fe09d512d960d9ae418f0aa9960ef16b40c2dad366048", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4e74a56a1c62e3", "s193a70d415430c", "s8b30a2a892ecd3", "saf8cba69a5faa2", "sc2859d1d29b588", "sd6e90f5e06563d", "s78652343516e3f" ] }, { "age": 80, "six": 47.12083333333333, "five": 49.218, "tasks": { "blimp_filtered": 67.14, "supplement_filtered": 57.01, "ewok_filtered": 50.51, "entity_tracking": 19.82, "comps": 51.61, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 49.0 }, "checkpoint": "8ce3fef475f407171b42861eaa935eb0dd2747ff85e4293ca82bce58b64c5ee4", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4e74a56a1c62e3", "s193a70d415430c", "s15df50f1249d9a", "s3c9d951b2bc98a", "sc2859d1d29b588", "sd6e90f5e06563d", "s78652343516e3f" ] }, { "age": 90, "six": 46.0825, "five": 48.166, "tasks": { "blimp_filtered": 66.36, "supplement_filtered": 53.71, "ewok_filtered": 49.52, "entity_tracking": 19.98, "comps": 51.26, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 49.0 }, "checkpoint": "6612d66646a40224c08b8d97f3a482b4dc373605e2ca94e95999e311a4553647", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4e74a56a1c62e3", "s193a70d415430c", "s02cc7c072a26ae", "s93ebe3069fd3d8", "sc2859d1d29b588", "sd6e90f5e06563d", "s78652343516e3f" ] }, { "age": 100, "six": 46.979166666666664, "five": 48.654, "tasks": { "blimp_filtered": 67.45, "supplement_filtered": 52.94, "ewok_filtered": 51.04, "entity_tracking": 19.98, "comps": 51.86, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 51.0 }, "checkpoint": "43c2bffa6cbe0e81ddad301bb0afe1615a104be62e029c46c705e2d73fa9a402", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4e74a56a1c62e3", "s193a70d415430c", "s485daca713e981", "sd71694c414ea71", "sc2859d1d29b588", "sd6e90f5e06563d", "s78652343516e3f" ] } ], "peak": { "five": { "score": 49.218, "ages": [ 80 ], "rank": 180 }, "six": { "score": 47.12083333333333, "ages": [ 80 ], "rank": 174 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 147, "id": "run/IPRLM336_FFN1120_ADDITIVE_SEED137_20261016", "original_name": "IPRLM336_FFN1120_ADDITIVE_SEED137_20261016", "title": "336幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 336, "ffn": 1120, "depth": 9, "parts": { "width": 336, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.3333333333333335 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 336, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M parent320/additive320/additive336FFN1176 and siblingFFN variants; observed later peaks separate." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 89, "sources": [ "sd7e857e3a25cd5", "s72e7b7fe623e33", "sc61b5651cdb686" ], "points": [ { "age": 40.0, "six": 47.119166666666665, "five": 49.81, "tasks": { "blimp_filtered": 68.02, "supplement_filtered": 56.59, "ewok_filtered": 51.54, "entity_tracking": 20.88, "comps": 52.02, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 45.0 }, "checkpoint": "6dc1c17b41cae59a3703139a329d2d1c5a99da18056c4d47e79deac2068ffa88", "precision": "fp32", "verification": "primary", "sources": [ "s0686452950ea6b", "s591c4a96bc8579" ] } ], "peak": { "five": { "score": 49.81, "ages": [ 40.0 ], "rank": 161 }, "six": { "score": 47.119166666666665, "ages": [ 40.0 ], "rank": 175 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 148, "id": "run/IPRLM384_FFN_INPUT_PATH_LR_V2_SEED20260907_40M_20261020", "original_name": "IPRLM384_FFN_INPUT_PATH_LR_V2_SEED20260907_40M_20261020", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "微細化V2：FFN入力層別LR", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005333333333333334, 0.00046666666666666666, 0.00039999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "NewfineFFNinputV2+8same-ageLRconditions; fullinventory;40Mnotfinalpeak" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 107, "sources": [ "sbdb26bd9b4b21e", "sd88aea12788c30" ], "points": [ { "age": 40.0, "six": 47.115, "five": 49.602, "tasks": { "blimp_filtered": 68.34, "supplement_filtered": 56.95, "ewok_filtered": 51.1, "entity_tracking": 19.19, "comps": 52.43, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 48.0 }, "checkpoint": "f428b7170791a12aa4d93676a014ca457bf1ed9017adb418207c10d2a4d97fbc", "precision": "fp32", "verification": "primary", "sources": [ "s2d6dd760dea217", "s41fe3c721478d9" ] } ], "peak": { "five": { "score": 49.602, "ages": [ 40.0 ], "rank": 170 }, "six": { "score": 47.115, "ages": [ 40.0 ], "rank": 176 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 149, "id": "run/IPRLM384_FFN1408_LRPLUS10_MAJOR110_SEED20260907_40M_20261020", "original_name": "IPRLM384_FFN1408_LRPLUS10_MAJOR110_SEED20260907_40M_20261020", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要103行列だけ1割増", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.0005866666666666668, 0.0005133333333333334, 0.00044 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "aftereachcase40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Three LR allocation cases and frozen baseline40M; full inventory; no final ranking from40M alone" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 116, "sources": [ "s1cc669b6ad33c4", "se9ce0be6afccaa" ], "points": [ { "age": 40.0, "six": 47.041666666666664, "five": 49.22, "tasks": { "blimp_filtered": 65.49, "supplement_filtered": 56.62, "ewok_filtered": 51.83, "entity_tracking": 20.19, "comps": 51.97, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 49.0 }, "checkpoint": "8b980fe3d718163c2cc737793638891aa4dcb878abb61ebd203348e7b94d0313", "precision": "fp32", "verification": "primary", "sources": [ "s5355459b717147", "s510b4afd995d81" ] } ], "peak": { "five": { "score": 49.22, "ages": [ 40.0 ], "rank": 179 }, "six": { "score": 47.041666666666664, "ages": [ 40.0 ], "rank": 177 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 150, "id": "baby/B_CYCLE_SYNC_LR_100M", "original_name": "系列長とLR同期巡回", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "local six-axis benchmark at every 10M checkpoint; no Reading or official submission evaluation", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 1104106, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "611332e27d15ac01cd3f65cce4c09b5c291fd0150732ed2c134cc21abf4ba121", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_ORDER_CONTENT_NO_OMISSIONS", "boundary_policy": "document-isolated causal masks; shuffled each pass; all prefixes and tails retained", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 4, "sources": [ "s8aa86092d38588", "sf5a646aaca0e91", "s71c2583a0c8370", "sd3eb2e8b715e94" ], "points": [ { "age": 10, "six": 45.700833333333335, "five": 47.908, "tasks": { "blimp_filtered": 66.2, "supplement_filtered": 56.91, "ewok_filtered": 49.29, "entity_tracking": 17.94, "comps": 49.2, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "0be9da27c6fce16e929373c67e6de9404629db8a16fde4e9eca9e11c0a1cad31", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s356adc3e69aed4", "s945205f5ddcfb5", "s265f688ed9fee9", "s8063f2834d3dbe", "sa170d6ca0570cc", "sbb1adb16a18a5c", "sce0f48d8f827b7" ] }, { "age": 20, "six": 45.77166666666667, "five": 48.384, "tasks": { "blimp_filtered": 67.98, "supplement_filtered": 56.2, "ewok_filtered": 48.75, "entity_tracking": 18.45, "comps": 50.54, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 46.0 }, "checkpoint": "2b0b0230990e7f8185790e7261224373d8bc08be92fe68d8ae6b983ceab46689", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s356adc3e69aed4", "s945205f5ddcfb5", "s95ad9670ed652a", "s7c93c45cb55765", "sa170d6ca0570cc", "sbb1adb16a18a5c", "sce0f48d8f827b7" ] }, { "age": 30, "six": 46.4075, "five": 49.156, "tasks": { "blimp_filtered": 67.34, "supplement_filtered": 58.29, "ewok_filtered": 50.49, "entity_tracking": 18.73, "comps": 50.93, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 43.0 }, "checkpoint": "a851fe44aec4f1c2d433689133440e4ad0ba919a055af27f9aa72848495a22b0", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s356adc3e69aed4", "s945205f5ddcfb5", "sda86fe7288085d", "se459f67e6495c5", "sa170d6ca0570cc", "sbb1adb16a18a5c", "sce0f48d8f827b7" ] }, { "age": 40, "six": 46.670833333333334, "five": 48.972, "tasks": { "blimp_filtered": 68.02, "supplement_filtered": 57.05, "ewok_filtered": 50.15, "entity_tracking": 18.52, "comps": 51.12, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 48.0 }, "checkpoint": "d7534be860c09a4afd81f78fc75143bf90489859a3d703b4b392f413b4f181b7", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s356adc3e69aed4", "s945205f5ddcfb5", "s0fde7fcaaca799", "sb38ae06c260b8b", "sa170d6ca0570cc", "sbb1adb16a18a5c", "sce0f48d8f827b7" ] }, { "age": 50, "six": 46.034166666666664, "five": 48.514, "tasks": { "blimp_filtered": 69.12, "supplement_filtered": 55.09, "ewok_filtered": 49.13, "entity_tracking": 18.08, "comps": 51.15, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 43.0 }, "checkpoint": "b4e17e0344362ee2ba78f363dcdf2a680dd532a6d281f17a1919b0866ffec8ab", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s356adc3e69aed4", "s945205f5ddcfb5", "s0a79818777099d", "s26395e16cbb50b", "sa170d6ca0570cc", "sbb1adb16a18a5c", "sce0f48d8f827b7" ] }, { "age": 60, "six": 45.681666666666665, "five": 48.382, "tasks": { "blimp_filtered": 68.89, "supplement_filtered": 54.82, "ewok_filtered": 49.88, "entity_tracking": 17.33, "comps": 50.99, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 43.0 }, "checkpoint": "f34b7e76745f176292ab2b3135b6d16f7e5484bfceae7cd43134c2e5eb4396e4", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s356adc3e69aed4", "s945205f5ddcfb5", "s090e0159b9bfc3", "s7ba9ae6e70c84c", "sa170d6ca0570cc", "sbb1adb16a18a5c", "sce0f48d8f827b7" ] }, { "age": 70, "six": 47.02916666666667, "five": 48.908, "tasks": { "blimp_filtered": 68.02, "supplement_filtered": 56.67, "ewok_filtered": 51.16, "entity_tracking": 17.8, "comps": 50.89, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 51.0 }, "checkpoint": "8a187bb0279704db449b0b68c20276b301fa0a2a040c5f086011c7e464d565d0", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s356adc3e69aed4", "s945205f5ddcfb5", "s5a7744c997bdae", "s34a9d4dab615e1", "sa170d6ca0570cc", "sbb1adb16a18a5c", "sce0f48d8f827b7" ] }, { "age": 80, "six": 46.865833333333335, "five": 48.912, "tasks": { "blimp_filtered": 68.8, "supplement_filtered": 56.21, "ewok_filtered": 49.95, "entity_tracking": 18.53, "comps": 51.07, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 49.0 }, "checkpoint": "85966cdd5903ac5fe7f860d776ed9c69de4875eb1b8627472da46b6f612fbc58", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s356adc3e69aed4", "s945205f5ddcfb5", "s76e5fd37bd442f", "s972d46d096b2f4", "sa170d6ca0570cc", "sbb1adb16a18a5c", "sce0f48d8f827b7" ] }, { "age": 90, "six": 46.479166666666664, "five": 48.942, "tasks": { "blimp_filtered": 68.57, "supplement_filtered": 56.44, "ewok_filtered": 50.53, "entity_tracking": 18.16, "comps": 51.01, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "379850690fa0da3a68140f26461e09ca77d02b119d6f859e7f2db6a0ef3e04f7", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s356adc3e69aed4", "s945205f5ddcfb5", "s43ef682d2d02ee", "sf664c60dd26e52", "sa170d6ca0570cc", "sbb1adb16a18a5c", "sce0f48d8f827b7" ] }, { "age": 100, "six": 46.041666666666664, "five": 48.714, "tasks": { "blimp_filtered": 68.06, "supplement_filtered": 57.01, "ewok_filtered": 49.58, "entity_tracking": 18.03, "comps": 50.89, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 44.0 }, "checkpoint": "293872364617a3dba161428e9665efd7248bbef455881d0cf9ed6b584df1d326", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s356adc3e69aed4", "s945205f5ddcfb5", "s94372577a4cd5b", "s0d5e896851f8ae", "sa170d6ca0570cc", "sbb1adb16a18a5c", "sce0f48d8f827b7" ] } ], "peak": { "five": { "score": 49.156, "ages": [ 30 ], "rank": 181 }, "six": { "score": 47.02916666666667, "ages": [ 70 ], "rank": 178 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 151, "id": "run/IPRLM_MOS3_256_LAST3_SEED731946281_40M_20261013", "original_name": "IPRLM_MOS3_256_LAST3_SEED731946281_40M_20261013", "title": "256幅・FFN896・9層", "topology": "MoS3", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last3", "attention_received_gate": true, "output_components": 3, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "After40M training: evaluate40M only", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.715, "comparison_scope": "Same parent320 architecture/LR/data order; one alternate seed; common70/80/90/100M peak and individual tasks. Two runs do not estimate seed variance precisely." }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 157, "sources": [ "seb7ab99a811f6c", "s5b96a006ed0dba", "s500886422d0429" ], "points": [ { "age": 40.0, "six": 46.973333333333336, "five": 49.818, "tasks": { "blimp_filtered": 68.73, "supplement_filtered": 57.7, "ewok_filtered": 51.88, "entity_tracking": 19.41, "comps": 51.37, "global_piqa_parallel": 16.5, "global_piqa_nonparallel": 49.0 }, "checkpoint": "56560f2efed35dbc791f181a1599b641a2f20e084310ef2123ff02a4fa43292c", "precision": "fp32", "verification": "primary", "sources": [ "s8d6c2c13b1f7de", "s9cfcc5bc85cd68" ] } ], "peak": { "five": { "score": 49.818, "ages": [ 40.0 ], "rank": 160 }, "six": { "score": 46.973333333333336, "ages": [ 40.0 ], "rank": 179 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 152, "id": "run/VISTA336_CANDIDATE_READ_CSR_FP32_SEED20260907_20261017", "original_name": "VISTA336_CANDIDATE_READ_CSR_FP32_SEED20260907_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": "未確認", "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "採点済み", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 158, "sources": [ "s99077488a7625d", "sf7a3c6427006d7", "se7876d3baebc83" ], "points": [ { "age": 40.0, "six": 46.93333333333333, "five": 49.878, "tasks": { "blimp_filtered": 68.06, "supplement_filtered": 57.29, "ewok_filtered": 51.61, "entity_tracking": 20.6, "comps": 51.83, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 45.0 }, "checkpoint": "af21a79cc96f79ec30cc90cc71e905e9d4c130218dcafffd6fbf8ba83c2086cc", "precision": "fp32", "verification": "primary", "sources": [ "sedc6481139e7fd", "sb1f60a349b134e" ] } ], "peak": { "five": { "score": 49.878, "ages": [ 40.0 ], "rank": 157 }, "six": { "score": 46.93333333333333, "ages": [ 40.0 ], "rank": 180 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 153, "id": "run/IPRLM_LR070605_320_SEED731946281_40M_20261013", "original_name": "IPRLM_LR070605_320_SEED731946281_40M_20261013", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "After40M training: evaluate40M only", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.715, "comparison_scope": "Same parent320 architecture/LR/data order; one alternate seed; common70/80/90/100M peak and individual tasks. Two runs do not estimate seed variance precisely." }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 156, "sources": [ "sf8021773cafd56", "sc01a73ab8bfb89", "s1c4853db597fb8" ], "points": [ { "age": 40.0, "six": 46.84583333333333, "five": 49.576, "tasks": { "blimp_filtered": 68.37, "supplement_filtered": 58.02, "ewok_filtered": 49.97, "entity_tracking": 19.18, "comps": 52.34, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "176e957044e2a720a7ab0ec2395ccd16cf008761d20f108cfcb4e1b4ec6cc3b3", "precision": "fp32", "verification": "primary", "sources": [ "s4d53297f036c75", "s449e6559225546" ] } ], "peak": { "five": { "score": 49.576, "ages": [ 40.0 ], "rank": 173 }, "six": { "score": 46.84583333333333, "ages": [ 40.0 ], "rank": 181 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 154, "id": "baby/B_READER_BATCH2048_BETA2_100M", "original_name": "旧Reader・batch2048＋β₂調整", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "Save checkpoints each10M; score all only after100M; local six axes; no official submission evaluation", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact FP32 eager CUDA graphs; TF32 disabled; choose validated score batching and graph chunk", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768 } }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 10, "sources": [ "sefad91db209ebf", "s3b5fa5f0996eee", "sa08127be39e829", "sf0a6b0945e8a70" ], "points": [ { "age": 10, "six": 45.12, "five": 47.608, "tasks": { "blimp_filtered": 63.19, "supplement_filtered": 55.22, "ewok_filtered": 49.73, "entity_tracking": 18.49, "comps": 51.41, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 44.0 }, "checkpoint": "c1c0ce5ecd7cb745733d40f9ccd17cdd95d47b2d17b84c3d80904e4c0785beef", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sccc2174b07f4a6", "s0da92175068aa4", "s770ca09b8e4947", "s222362390d8e95", "s92beb8f86bf795", "s9cad431a9bcced", "s41782f114cea34", "sa76eaf5e04adb0", "sdb77647392fe52" ] }, { "age": 20, "six": 43.07, "five": 44.748, "tasks": { "blimp_filtered": 55.19, "supplement_filtered": 51.82, "ewok_filtered": 49.68, "entity_tracking": 16.62, "comps": 50.43, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 48.0 }, "checkpoint": "296a94238edc61d2555ed248ca602c5e06ac1e3a3551b1dad164bfecea28539f", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sccc2174b07f4a6", "s0da92175068aa4", "s0ce161fd229a9e", "sd93effc9e9a483", "s41782f114cea34", "sa76eaf5e04adb0", "sdb77647392fe52" ] }, { "age": 30, "six": 43.76583333333333, "five": 45.198, "tasks": { "blimp_filtered": 55.77, "supplement_filtered": 54.09, "ewok_filtered": 49.28, "entity_tracking": 17.13, "comps": 49.72, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 47.0 }, "checkpoint": "b0ff27e9322bf55b7e74137439307710b422ed112eb2de63e213bdb6c730a28a", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sccc2174b07f4a6", "s0da92175068aa4", "s55b9de316cd8e6", "sadefe0e88e14e9", "s41782f114cea34", "sa76eaf5e04adb0", "sdb77647392fe52" ] }, { "age": 40, "six": 44.0725, "five": 45.36, "tasks": { "blimp_filtered": 57.62, "supplement_filtered": 52.77, "ewok_filtered": 49.51, "entity_tracking": 16.89, "comps": 50.01, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 51.0 }, "checkpoint": "b63031c48ceba2186f4616e1cca897926959384f7d93657488436564024ee1a1", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sccc2174b07f4a6", "s0da92175068aa4", "s28edfa142b85b8", "s17406ec5c896dd", "s41782f114cea34", "sa76eaf5e04adb0", "sdb77647392fe52" ] }, { "age": 50, "six": 44.388333333333335, "five": 46.33, "tasks": { "blimp_filtered": 57.83, "supplement_filtered": 56.19, "ewok_filtered": 49.86, "entity_tracking": 17.23, "comps": 50.54, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 48.0 }, "checkpoint": "2d0f73920ed2e50cd6dc1332a83c7ff6087e61b6623eaa1fdcbfa57d62f13fb1", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sccc2174b07f4a6", "s0da92175068aa4", "s1ba7ce9c86f15b", "s59f3c40ac271ce", "s41782f114cea34", "sa76eaf5e04adb0", "sdb77647392fe52" ] }, { "age": 60, "six": 43.91583333333333, "five": 45.472, "tasks": { "blimp_filtered": 58.92, "supplement_filtered": 51.97, "ewok_filtered": 49.69, "entity_tracking": 16.96, "comps": 49.82, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 48.0 }, "checkpoint": "cf6067eebc663296d52b254c4c126f6fade4aebf377dbd46bc3b25275dc88aa0", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sccc2174b07f4a6", "s0da92175068aa4", "s5a090b74d93114", "sd792f68a3e6a29", "s41782f114cea34", "sa76eaf5e04adb0", "sdb77647392fe52" ] }, { "age": 70, "six": 44.9875, "five": 46.852, "tasks": { "blimp_filtered": 60.22, "supplement_filtered": 56.17, "ewok_filtered": 49.56, "entity_tracking": 17.59, "comps": 50.72, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 49.0 }, "checkpoint": "f2054a9776cb05e11c325b5282767b30e0328fa1d92a8828b9e43f144443032e", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sccc2174b07f4a6", "s0da92175068aa4", "sa3b6360c573e31", "s9f8d2e2b7c7215", "s41782f114cea34", "sa76eaf5e04adb0", "sdb77647392fe52" ] }, { "age": 80, "six": 44.245, "five": 46.358, "tasks": { "blimp_filtered": 58.51, "supplement_filtered": 56.77, "ewok_filtered": 49.0, "entity_tracking": 17.21, "comps": 50.3, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 46.0 }, "checkpoint": "017fb69682af87a83bc04840a864a38adfa9961eb7fc1caeca087ff6783476e6", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sccc2174b07f4a6", "s0da92175068aa4", "s9d2fb63ffdf83b", "s568704859c276d", "s41782f114cea34", "sa76eaf5e04adb0", "sdb77647392fe52" ] }, { "age": 90, "six": 44.715, "five": 46.322, "tasks": { "blimp_filtered": 60.74, "supplement_filtered": 53.73, "ewok_filtered": 49.72, "entity_tracking": 17.0, "comps": 50.42, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 52.0 }, "checkpoint": "4ac5c059d8db213a6a393dfde43fe84d322058efe0ae97fbf062bac22a69c147", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sccc2174b07f4a6", "s0da92175068aa4", "sf2934dfcb73a0e", "s40d62b0fc62006", "s41782f114cea34", "sa76eaf5e04adb0", "sdb77647392fe52" ] }, { "age": 100, "six": 44.75833333333333, "five": 47.062, "tasks": { "blimp_filtered": 60.44, "supplement_filtered": 56.26, "ewok_filtered": 49.09, "entity_tracking": 18.77, "comps": 50.75, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 49.0 }, "checkpoint": "7320f340b65f6ec86e9a2761ccf2026fc2d4a2b3fdb8731195eabb98658dcddf", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sccc2174b07f4a6", "s0da92175068aa4", "s0f877b9d6a7b16", "sb5ffa3e4112973", "s41782f114cea34", "sa76eaf5e04adb0", "sdb77647392fe52" ] } ], "peak": { "five": { "score": 47.608, "ages": [ 10 ], "rank": 184 }, "six": { "score": 45.12, "ages": [ 10 ], "rank": 183 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 155, "id": "baby/B_CONTEXT_SPEED_REVIEW_20260930", "original_name": "B_CONTEXT_SPEED_REVIEW_20260930", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "no_scored_checkpoint_in_snapshot", "architecture": { "id": "VISTARC", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "first_pass_loss_weight": 0.5 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "local six-axis benchmark at every 10M checkpoint; no Reading or official submission evaluation", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "retain_backward_graph": true, "compile_output_tiles": false, "retained_backward_policy": "Eager Transformer body and output tiles: full packed batch retained backward corrupts compiled body gradients. Preserve vectorized fused auxiliary head and CUDA fused AdamW.", "compile_hidden": false, "output_head_memory_policy": "8192-position tiles and non-reentrant checkpoint recomputation for both first-pass feedback/CE and final CE", "body_memory_policy": "Non-reentrant block checkpoint at sequence length512, preserve dropout RNG; no block checkpoint at128/256" }, "condition_conflicts": [], "profile": 2, "sources": [ "s40c7844cf55453" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 156, "id": "baby/B_LR_COSINE_ABORTED_END1E4_PREFIX_20260929", "original_name": "B_LR_COSINE_ABORTED_END1E4_PREFIX_20260929", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "旧記録RUNNING・現状未確認", "architecture": { "id": "VISTARC", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "local six-axis benchmark at every 10M checkpoint; no Reading or official submission evaluation", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 1104106, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "611332e27d15ac01cd3f65cce4c09b5c291fd0150732ed2c134cc21abf4ba121", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_ORDER_CONTENT_NO_OMISSIONS", "boundary_policy": "document-isolated causal masks; shuffled each pass; all prefixes and tails retained", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 65, "sources": [ "s678ff7e69dbb43" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 157, "id": "baby/B_READER_DEPTH9_DISTANCE_LOG_IMPLEMENTATION", "original_name": "B_READER_DEPTH9_DISTANCE_LOG_IMPLEMENTATION", "title": "256幅・FFN896・9層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "学習する距離補正", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "no_scored_checkpoint_in_snapshot", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_distance": "learned_log", "distance_init_min": 0.0625, "distance_init_max": 1.0 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 16, "sources": [ "s19f490a160ac89", "s946ab650b6d18c", "s61ec17f52202df", "s7ff8522ab2e007" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 158, "id": "baby/B_READER_DEPTH9_WIDTH256_NO_DISTANCE_TWO_KNOBS_100M", "original_name": "B_READER_DEPTH9_WIDTH256_NO_DISTANCE_TWO_KNOBS_100M", "title": "256幅・FFN896・9層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "no_scored_checkpoint_in_snapshot", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 66, "sources": [ "s13ea1864fc30df", "s1ab947bc15c298", "sfc9217ffbe8220", "se004c774c9e8e3" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 159, "id": "baby/B_READER_DEPTH9_WIDTH256_QKNORM_GAIN_SQRT32_LR1_100M", "original_name": "B_READER_DEPTH9_WIDTH256_QKNORM_GAIN_SQRT32_LR1_100M", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "no_scored_checkpoint_in_snapshot", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 67, "sources": [ "sdf0db2438e9d8f", "sb155dc8c091256", "s08d98ea9097dff", "s148e9330636489" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 160, "id": "baby/VISTARC_READER_REPAIRED_V1", "original_name": "VISTARC_READER_REPAIRED_V1", "title": "256幅・FFN896・8層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 8, "parts": { "width": 256, "ffn": 896, "depth": 8, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "no_scored_checkpoint_in_snapshot", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 8, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "vocab": 16384, "seed": 20260907, "reader_feature_mode": "bounded_innovation", "innovation_variance_floor": 0.1, "spread_direction_floor": 0.01 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "Save each10M without stopping training; score80M and100M after100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact FP32 eager CUDA graphs; TF32 disabled; choose validated score batching and graph chunk", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "checkpoint_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": false, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "gpu_power_limit_watts": 260 }, "condition_conflicts": [], "profile": 68, "sources": [ "sa85c0f44aa42da", "s53ade80c3fec2e", "sb71f91bf10b663", "sa707c6f460e89b" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 161, "id": "baby/VISTA_STATE_LOOP_2_4X3_2_100M", "original_name": "VISTA_STATE_LOOP_2_4X3_2_100M", "title": "256幅・FFN896・8層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 8, "parts": { "width": 256, "ffn": 896, "depth": 8, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "no_scored_checkpoint_in_snapshot", "architecture": { "id": "VISTA_STATE_LOOP", "width": 256, "layers": 8, "hidden": 896, "attention_heads": 8, "head_dim": 32, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "knob_lr_ratio": 0.3, "dropout_init": 0.0, "vocab": 16384, "seed": 20260907, "prelude_layers": 2, "core_layers": 4, "coda_layers": 2, "loops": 3, "effective_depth": 16, "state_adapter_init": [ 0.9, 0.1 ], "residual_init_effective_depth": 16 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score80M and100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "IEEE FP32; parallel causal forward with streamed16-cache and padding checks; canonical stream fallback on failure; unchanged score tolerances", "runtime_selection": { "max_batch": 128, "token_budget": 8192, "mode": "parallel_fp32_validated_against16cache" }, "benchmark_words": [ 80000000, 100000000 ], "speed_preflight": "Same448 real questions,128 versus256 row batching, all scores and choices checked against independent16-cache streaming; no benchmark-based hyperparameter selection" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": false, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260, "residual_states_and_delta_history": "fp32" }, "condition_conflicts": [], "profile": 69, "sources": [ "se81bbbb750263b", "s3a5f773984f289", "sffe44efb7732fd", "s1c1d09fe4c1c74" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 162, "id": "baby/checkpoint_fix", "original_name": "checkpoint_fix", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "no_scored_checkpoint_in_snapshot", "architecture": { "id": "VISTARC", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "local six-axis benchmark at every 10M checkpoint; no Reading or official submission evaluation", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 1104106, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "611332e27d15ac01cd3f65cce4c09b5c291fd0150732ed2c134cc21abf4ba121", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_ORDER_CONTENT_NO_OMISSIONS", "boundary_policy": "document-isolated causal masks; shuffled each pass; all prefixes and tails retained", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 3, "sources": [ "s111b09b8e4b43b" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 163, "id": "baby/vistarc_q0_content_selective_qv_20261004", "original_name": "Q0・内容選別QV", "title": "256幅・FFN896・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 9, "parts": { "width": 256, "ffn": 896, "depth": 9, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 9, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "none", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "content_selective_qv": true, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local five axes only" ], "intermediate": "After100M only,60/70/80/90/100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ], "suite": "local five axes, PIQA excluded", "tasks": [ [ "blimp", "blimp_filtered", 59875 ], [ "blimp", "supplement_filtered", 5218 ], [ "ewok", "ewok_filtered", 7618 ], [ "entity_tracking", "entity_tracking", 6780 ], [ "comps", "comps", 91028 ] ], "objective": "highest observed five-axis checkpoint; threshold51.014; no penalty for later decline" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 39, "sources": [ "s73d9453501797d", "s0d507599eb01a3", "sf2c20941ab55f5", "s20f3f918162e06", "s4d4ceb7b797700" ], "points": [ { "age": 60, "six": null, "five": 49.728, "tasks": { "blimp_filtered": 68.35, "supplement_filtered": 57.04, "ewok_filtered": 50.78, "entity_tracking": 20.34, "comps": 52.13 }, "checkpoint": "05f26d64cb565650046aa7b049924211eb563ea1f1256599e13bbf77022290f1", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s4815ce5be28a83", "s592bf1ee098883", "s81a21dbc045602", "s35fb4c3fb768ec", "sa0404434278362", "s432eee017ae2cb", "sfbe404db2eb725" ] }, { "age": 70, "six": null, "five": 50.52, "tasks": { "blimp_filtered": 68.4, "supplement_filtered": 59.8, "ewok_filtered": 51.44, "entity_tracking": 20.61, "comps": 52.35 }, "checkpoint": "e9d26f4b0d1572e27324ecdddf322fc78690f924a4b681204d9f700ce0e09a21", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s2fc4d52374bb61", "s592bf1ee098883", "s1f546d3a9027e5", "s35fb4c3fb768ec", "sa0404434278362", "s432eee017ae2cb", "sfbe404db2eb725" ] }, { "age": 80, "six": null, "five": 50.582, "tasks": { "blimp_filtered": 68.92, "supplement_filtered": 61.16, "ewok_filtered": 51.41, "entity_tracking": 19.05, "comps": 52.37 }, "checkpoint": "633f1d6584c565b48ef41b5b80c457386c497ed91b646986efbf55e7e308d16a", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s87ec263edc6e05", "s592bf1ee098883", "s974d8ba68b12b8", "s35fb4c3fb768ec", "sa0404434278362", "s432eee017ae2cb", "sfbe404db2eb725" ] }, { "age": 90, "six": null, "five": 49.474, "tasks": { "blimp_filtered": 68.89, "supplement_filtered": 57.44, "ewok_filtered": 50.3, "entity_tracking": 18.53, "comps": 52.21 }, "checkpoint": "eba070c5463410c071640fe856870b992002ed62c115433fe51a98c74311b88c", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s9b9a857616c093", "s592bf1ee098883", "s3634bb0f992186", "s35fb4c3fb768ec", "sa0404434278362", "s432eee017ae2cb", "sfbe404db2eb725" ] }, { "age": 100, "six": null, "five": 50.44, "tasks": { "blimp_filtered": 69.14, "supplement_filtered": 58.97, "ewok_filtered": 51.51, "entity_tracking": 20.23, "comps": 52.35 }, "checkpoint": "116892a932f007fbb37d1916814623464e01deae4fbbc332a76a2a0a21798af5", "precision": "fp32", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s6e8dcc4cab1862", "s592bf1ee098883", "se92857850eebc9", "s35fb4c3fb768ec", "sa0404434278362", "s432eee017ae2cb", "sfbe404db2eb725" ] } ], "peak": { "five": { "score": 50.582, "ages": [ 80 ], "rank": 103 }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 164, "id": "run/B_READER_DEPTH9_WIDTH272_QKNORM_HEAD_GAIN_100M", "original_name": "B_READER_DEPTH9_WIDTH272_QKNORM_HEAD_GAIN_100M", "title": "272幅・FFN960・9層", "topology": "単一出力", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 272, "ffn": 960, "depth": 9, "parts": { "width": 272, "ffn": 960, "depth": 9, "attention": "8×34", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "なし", "output": "単一出力", "reference": "未確認", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 1, "reference_mode": "未確認", "ffn_ratio": 3.5294117647058822 }, "main_lr": [], "state": "評価記録なし", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 272, "layers": 9, "hidden": 960, "attention_heads": 8, "head_dim": 34, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 32, "history_norm": "separate", "seed": 20260907, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "attention_width": 272, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "attention_received_gate": false, "ffn_input_output_gates": false, "output_head": "single", "output_components": 1 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No intermediate evaluation; score60/70/80/90/100M after100M training", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Canonical prefix FP32 eager CUDA graphs; TF32 disabled; unchanged strict tolerances and canonical fallback", "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "benchmark_words": [ 80000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 169, "sources": [ "sb4042f89e9bfdd" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 165, "id": "run/IPRLM320_FFN1200_ADDITIVE_SEED20260907_20261019", "original_name": "IPRLM320_FFN1200_ADDITIVE_SEED20260907_20261019", "title": "320幅・FFN1200・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1200, "depth": 9, "parts": { "width": 320, "ffn": 1200, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.75 }, "main_lr": [], "state": "CANCELLED_BY_USER", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1200, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at70/80/90/100M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Inherited exact FP32 recurrent prefix graph; new model GPU parity and capacity unmeasured", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "Inherited future template only; execution not authorized", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Full known inventory plus matched-age parent320; observed peaks retained separately" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ], "execution_authorized": false, "scope_note": "Inherited intended70/80/90/100M six-axis template; execution requires later user GO." }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 75, "sources": [ "s6938e3c8fb499e", "s088075b1f4f9d0", "s1988462957b2a9" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 166, "id": "run/IPRLM320_FFN_CONTEXTUAL_ADDITION_SEED137_20261019", "original_name": "IPRLM320_FFN_CONTEXTUAL_ADDITION_SEED137_20261019", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "評価記録なし", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true, "ffn_contextual_bypass": true, "ffn_context_scale": 0.003125 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "proposed local six axes at40M; not launched" ], "intermediate": "None executed; launch requires a later instruction.", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "proposed after each40M run", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Matched parent, unconditional and conditional addition at40M; no final peak claim from40M alone" }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 80, "sources": [ "s76120ce29d090c", "sbea2b9634f46e4" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 167, "id": "run/IPRLM320_FFN_CONTEXTUAL_ADDITION_SEED731946281_20261019", "original_name": "IPRLM320_FFN_CONTEXTUAL_ADDITION_SEED731946281_20261019", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "評価記録なし", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true, "ffn_contextual_bypass": true, "ffn_context_scale": 0.003125 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "proposed local six axes at40M; not launched" ], "intermediate": "None executed; launch requires a later instruction.", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "proposed after each40M run", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Matched parent, unconditional and conditional addition at40M; no final peak claim from40M alone" }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 80, "sources": [ "sd80a64b12e5c03", "s44f3180eba1a31" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 168, "id": "run/IPRLM320_GATE_SHORT_SEARCH_20261011", "original_name": "IPRLM320_GATE_SHORT_SEARCH_20261011", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "prediction_residual_v1", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "評価記録なし", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_init_only", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123_prediction_gate", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "initialization_layers": 9, "output_mixture_gate": "prediction_residual_v1", "mixture_gate_hidden": 320, "mixture_gate_statistics": "normalized_entropy_js_v1", "ffn_input_output_gates": false }, "evaluation": { "mode": "fixed640 local diagnostic", "per_task": 128, "tasks": [ "blimp_filtered", "supplement_filtered", "ewok_filtered", "entity_tracking", "comps" ], "precision": "fp32_tf32off_exact_recurrence", "official_rank_claim": false, "full_benchmarks": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 85, "sources": [ "s0b79c6f66b2508", "sbc9c92f98b8e17", "sdd92f696d52013" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 169, "id": "run/IPRLM336_FFN1288_ADDITIVE_SEED137_20261017", "original_name": "IPRLM336_FFN1288_ADDITIVE_SEED137_20261017", "title": "336幅・FFN1288・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 336, "ffn": 1288, "depth": 9, "parts": { "width": 336, "ffn": 1288, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "あり", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.8333333333333335 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 336, "layers": 9, "hidden": 1288, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M parent320, additive320,336FFN1120/1176/1232/1288; later peaks separate." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 92, "sources": [ "s53d99492db2a89", "s2896794eb7bfbf", "sb709e651123ed3" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 170, "id": "run/IPRLM384_B_INPUT820_VOCAB504_SEED20260907_40M_20261021", "original_name": "IPRLM384_B_INPUT820_VOCAB504_SEED20260907_40M_20261021", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00082", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "STOPPED_BY_USER", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All20same-seed40M LRconditions;all7tasks,5/6aggregates,pairedoutcomes;fullinventory" }, "milestones_words": [ 40000000 ], "checkpoints_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 102, "sources": [ "sb410fbe672ab3b", "s28b687be3e82b0" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 171, "id": "run/Q3_MOS2_LAST3_WIDTH384_20261009", "original_name": "Q3_MOS2_LAST3_WIDTH384_20261009", "title": "384幅・FFN1344・9層", "topology": "MoS2・共通3層混合", "recipe": "主要0.0005333／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1344, "depth": 9, "parts": { "width": 384, "ffn": 1344, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS2", "reference": "7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 2, "reference_mode": "共通で混合", "reference_layers": [ [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.0005333333333333334, 0.00046666666666666666, 0.00039999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1344, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos2_last3", "attention_received_gate": true, "output_components": 2, "output_state_layers": 3, "reference_attention_width": 256, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym primary31 suites", "original PIQA validation1838" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 exact recurrent prefix graph; checkpoint matched MoS2 last3; TF32off; no changes to probability scoring", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": [ "five axes without PIQA", "SyntaxGym31", "original PIQA acc/acc_norm" ], "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.715, "comparison_scope": "256MoS2/3/4 last3 and384MoS3 last2; external fixed80M peers; own50-100M trajectory" }, "milestones_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "external_protocol": "source:sd1eef500b3fff3" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 165, "sources": [ "saaa0135659bc29", "s85d908d4b5b072", "s465ab54f6fd674", "s2b0d9de143204f" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 172, "id": "run/Q3_MOS3_BRANCH122_WIDTH384_FFN1344_HEADS12_20261009", "original_name": "Q3_MOS3_BRANCH122_WIDTH384_FFN1344_HEADS12_20261009", "title": "384幅・FFN1344・9層", "topology": "MoS3・枝別1/2/2層混合", "recipe": "主要0.0005333／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1344, "depth": 9, "parts": { "width": 384, "ffn": 1344, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.0005333333333333334, 0.00046666666666666666, 0.00039999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1344, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch122", "attention_received_gate": true, "output_components": 3, "output_state_layers": 2, "reference_attention_width": 256, "output_reference_counts": [ 1, 2, 2 ], "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 8, 9 ] ], "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Inherited runtime candidate only; Branch122 GPU parity/speed/VRAM acceptance pending", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 65536, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.138, "same_family_peak_six_axis": 49.715, "comparison_scope": "384_last2 parent,384_last1,256_last2; compare matching evaluated checkpoints and observed peaks", "parent_peak_six_axis": 49.715, "same_family_peak_five_axis": 51.72 }, "milestones_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "evaluation_authorized": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 166, "sources": [ "sb1fbb81065d886", "s86f9cc44408074", "sca5f1c422e6ffd", "s599fbed5b581e6" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 173, "id": "run/VISTA336_CANDIDATE_READ_CSR_FP32_SEED137_20261017", "original_name": "VISTA336_CANDIDATE_READ_CSR_FP32_SEED137_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": "未確認", "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 158, "sources": [ "s462e1cb9a5a129", "s8690f49594e13f", "sc085a16b10c176" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 174, "id": "run/VISTA336_CANDIDATE_READ_CSR_FP32_SEED731946281_20261017", "original_name": "VISTA336_CANDIDATE_READ_CSR_FP32_SEED731946281_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": "未確認", "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 158, "sources": [ "s77dbb4b8e9c419", "s6e3d6dee157ae3", "sea90a2117de533" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 175, "id": "run/VISTA336_CANDIDATE_READ_FULL_SEED137_20261017", "original_name": "VISTA336_CANDIDATE_READ_FULL_SEED137_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": "未確認", "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 158, "sources": [ "s4a1cc9264edd85", "seaf94bae3182cb", "sd92348e9e2e4f4" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 176, "id": "run/VISTA336_CANDIDATE_READ_FULL_SEED137_20261017_MASK_RUNTIME_REBUILD", "original_name": "VISTA336_CANDIDATE_READ_FULL_SEED137_20261017_MASK_RUNTIME_REBUILD", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": "未確認", "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 158, "sources": [ "s8ab8fc87f38215", "s1b1237df4cc644", "s526e2348f07422" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 177, "id": "run/VISTA336_CANDIDATE_READ_FULL_SEED20260907_20261017", "original_name": "VISTA336_CANDIDATE_READ_FULL_SEED20260907_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": "未確認", "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 158, "sources": [ "s17bb7a5cc205f0", "scf779f2dffff00", "s982557ce684917" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 178, "id": "run/VISTA336_CANDIDATE_READ_FULL_SEED20260907_20261017_MASK_RUNTIME_REBUILD", "original_name": "VISTA336_CANDIDATE_READ_FULL_SEED20260907_20261017_MASK_RUNTIME_REBUILD", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": "未確認", "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 158, "sources": [ "s85ecc008197214", "s9198f4cf516be0", "sb2227571fc84d0" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 179, "id": "run/VISTA336_CANDIDATE_READ_FULL_SEED731946281_20261017", "original_name": "VISTA336_CANDIDATE_READ_FULL_SEED731946281_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": "未確認", "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 158, "sources": [ "sc22fb691b8413a", "s3e9ed19570e0ef", "s34ba25413ffc9e" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 180, "id": "run/VISTA336_CANDIDATE_READ_FULL_SEED731946281_20261017_MASK_RUNTIME_REBUILD", "original_name": "VISTA336_CANDIDATE_READ_FULL_SEED731946281_20261017_MASK_RUNTIME_REBUILD", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": "未確認", "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "learn_ffn_variance": true, "attention_gate_bias": true, "candidate_support": "all_vocabulary_within_learned_delta", "candidate_read_width": 336 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 158, "sources": [ "sbf2debf80c34d0", "s991af9d159a3ba", "sa98132f489daaf" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 181, "id": "run/VISTA336_CANDIDATE_READ_SEED137_20261017", "original_name": "VISTA336_CANDIDATE_READ_SEED137_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 16, "candidate": "各候補を受信状態で再配分", "reader": 336, "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "なし", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 0.0, "router_width": 16, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 336, "attention_distance": "none", "learn_rope_base": false, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "candidate_count": 64, "learn_ffn_variance": false, "attention_gate_bias": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 164, "sources": [ "s1e4e08d7fb478a", "s9f6f93f52b2cf0", "s3a397c15cd2947" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 182, "id": "run/VISTA336_CANDIDATE_READ_SEED20260907_20261017", "original_name": "VISTA336_CANDIDATE_READ_SEED20260907_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 16, "candidate": "各候補を受信状態で再配分", "reader": 336, "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "なし", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 0.0, "router_width": 16, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 336, "attention_distance": "none", "learn_rope_base": false, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "candidate_count": 64, "learn_ffn_variance": false, "attention_gate_bias": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 164, "sources": [ "s3b756caa30f5f5", "scf18f1054c9e32", "sfe203ddc440f62" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 183, "id": "run/VISTA336_CANDIDATE_READ_SEED731946281_20261017", "original_name": "VISTA336_CANDIDATE_READ_SEED731946281_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 16, "candidate": "各候補を受信状態で再配分", "reader": 336, "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "なし", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 0.0, "router_width": 16, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 336, "attention_distance": "none", "learn_rope_base": false, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "candidate_count": 64, "learn_ffn_variance": false, "attention_gate_bias": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 164, "sources": [ "sf4c426c14e4ed3", "s8e3666fe0cc854", "s9e9678e42fd5ab" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 184, "id": "run/VISTA336_CANDIDATE_READ_TUNED_SEED137_20261017", "original_name": "VISTA336_CANDIDATE_READ_TUNED_SEED137_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": 336, "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 336, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 137, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "candidate_count": 64, "learn_ffn_variance": true, "attention_gate_bias": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 158, "sources": [ "s9ede7832625135", "sfb3ba693c9337b", "sb58f0a5bc96618" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 185, "id": "run/VISTA336_CANDIDATE_READ_TUNED_SEED20260907_20261017", "original_name": "VISTA336_CANDIDATE_READ_TUNED_SEED20260907_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": 336, "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 336, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "candidate_count": 64, "learn_ffn_variance": true, "attention_gate_bias": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 158, "sources": [ "s1a284c538825c9", "s41a8f69c1b32a0", "s31c5fd4a2cf084" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 186, "id": "run/VISTA336_CANDIDATE_READ_TUNED_SEED731946281_20261017", "original_name": "VISTA336_CANDIDATE_READ_TUNED_SEED731946281_20261017", "title": "336幅・FFN1232・9層", "topology": "単一出力・枝別1層混合", "recipe": "主要0.00064／Embedding0.0008", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 336, "ffn": 1232, "depth": 9, "parts": { "width": 336, "ffn": 1232, "depth": 9, "attention": "10×32", "router": 32, "candidate": "各候補を受信状態で再配分", "reader": 336, "received": "なし", "output": "単一出力", "reference": "9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00064, 0.0005600000000000001, 0.00047999999999999996 ], "state": "評価記録なし", "architecture": { "id": "VISTA_CANDIDATE_READ", "width": 336, "layers": 9, "hidden": 1232, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 336, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": false, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "single", "attention_received_gate": false, "output_components": 1, "output_state_layers": 1, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ] ], "ffn_additive_bypass": false, "candidate_count": 64, "learn_ffn_variance": true, "attention_gate_bias": true }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M" ], "intermediate": "After eachseed40M; all3seed comparison before extension", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "Exact recurrent FP32 single output; CUDA prefix graph verified against canonical stream", "runtime_selection": { "graph_chunk": 4, "max_batch": 2048, "kv_token_budget": 81920, "prefix_graph": true, "fused_mixture": false, "merge_blocks": 16 }, "benchmark_words": [ 40000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000 ], "timing": "after eachseed40M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Same-seed40M versus additive336FFN1232; whole-system comparison, not candidate-read-only causal effect." }, "milestones_words": [ 40000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 158, "sources": [ "s0dec98471fa71b", "se64fb848165f61", "sbd0e460dbe019f" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 187, "id": "run/VISTARC", "original_name": "VISTARC", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "評価記録なし", "architecture": { "id": "VISTARC", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "full zero-shot", "GlobalPIQA", "reading", "AoA", "SuperGLUE fine-tuning" ], "intermediate": "official fast suite at all 19 checkpoints", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 1104106, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "611332e27d15ac01cd3f65cce4c09b5c291fd0150732ed2c134cc21abf4ba121", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_ORDER_CONTENT_NO_OMISSIONS", "boundary_policy": "document-isolated causal masks; shuffled each pass; all prefixes and tails retained" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 3, "sources": [ "s5e34488fd315aa", "s9b25018cbc6018", "s661c0065007e41" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 188, "id": "run/VISTARC_SMALL_MAXPERF_V1", "original_name": "VISTARC_SMALL_MAXPERF_V1", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "評価記録なし", "architecture": { "id": "VISTARC", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.16, "tau_init": 1.15, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "delta_init": 3.0, "alpha_lr_ratio": 0.6, "tau_lr_ratio": 0.6, "delta_lr_ratio": 0.15 }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "full zero-shot", "GlobalPIQA", "reading", "AoA", "SuperGLUE fine-tuning" ], "intermediate": "official fast suite at all 19 checkpoints", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 1104106, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "9f726609d2613d90faba47344d7435e72c9b96fde41a272e36a6107935eeca3b", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_ORDER_CONTENT_NO_OMISSIONS", "boundary_policy": "document-isolated causal masks; shuffled each pass; all prefixes and tails retained" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 168, "sources": [ "sb9ac5f09343b49" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 189, "id": "legacy-dense/depth14", "original_name": "旧Dense・14層", "title": "旧Dense・14層", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.534166666666664, "five": 49.202, "tasks": { "blimp_filtered": 67.6, "supplement_filtered": 55.58, "ewok_filtered": 51.75, "entity_tracking": 19.78, "comps": 51.3, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scc0a962e01ae79" ] }, { "age": 30, "six": 46.44833333333333, "five": 48.908, "tasks": { "blimp_filtered": 67.36, "supplement_filtered": 56.04, "ewok_filtered": 51.0, "entity_tracking": 18.45, "comps": 51.69, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 45.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s28329bd7a2322f" ] }, { "age": 40, "six": 46.884166666666665, "five": 49.822, "tasks": { "blimp_filtered": 69.74, "supplement_filtered": 57.14, "ewok_filtered": 51.33, "entity_tracking": 19.05, "comps": 51.85, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 44.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scc0a962e01ae79" ] }, { "age": 50, "six": 46.79333333333334, "five": 49.116, "tasks": { "blimp_filtered": 68.32, "supplement_filtered": 56.45, "ewok_filtered": 51.01, "entity_tracking": 18.14, "comps": 51.66, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 49.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scc0a962e01ae79" ] }, { "age": 60, "six": 46.85916666666667, "five": 49.104, "tasks": { "blimp_filtered": 68.0, "supplement_filtered": 56.33, "ewok_filtered": 50.91, "entity_tracking": 18.5, "comps": 51.78, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s28329bd7a2322f" ] }, { "age": 70, "six": 47.8825, "five": 49.332, "tasks": { "blimp_filtered": 69.01, "supplement_filtered": 56.96, "ewok_filtered": 50.29, "entity_tracking": 18.62, "comps": 51.78, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 57.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scc0a962e01ae79" ] }, { "age": 80, "six": 48.10166666666667, "five": 49.592, "tasks": { "blimp_filtered": 69.14, "supplement_filtered": 57.81, "ewok_filtered": 49.36, "entity_tracking": 19.42, "comps": 52.23, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 58.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scc0a962e01ae79" ] }, { "age": 90, "six": 47.001666666666665, "five": 49.566, "tasks": { "blimp_filtered": 68.38, "supplement_filtered": 57.82, "ewok_filtered": 51.19, "entity_tracking": 18.33, "comps": 52.11, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scc0a962e01ae79" ] }, { "age": 100, "six": 47.208333333333336, "five": 49.42, "tasks": { "blimp_filtered": 68.9, "supplement_filtered": 56.65, "ewok_filtered": 50.84, "entity_tracking": 18.86, "comps": 51.85, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 49.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s28329bd7a2322f" ] } ], "peak": { "five": { "score": 49.822, "ages": [ 40 ], "rank": 2 }, "six": { "score": 48.10166666666667, "ages": [ 80 ], "rank": 1 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 190, "id": "legacy-dense/direct", "original_name": "通常VISTA・7層・分散全幅直接", "title": "通常VISTA・7層・分散全幅直接", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.53666666666667, "five": 48.014, "tasks": { "blimp_filtered": 65.33, "supplement_filtered": 57.34, "ewok_filtered": 49.8, "entity_tracking": 17.78, "comps": 49.82, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 43.0 }, "checkpoint": "eca7973f4410799ddfe9b84148a05f778ebb6d1bc4deaf1d4a3eb4ef5bad8651", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s82b5b7083b761e" ] }, { "age": 20, "six": 46.22416666666667, "five": 48.63, "tasks": { "blimp_filtered": 67.45, "supplement_filtered": 56.46, "ewok_filtered": 49.97, "entity_tracking": 18.26, "comps": 51.01, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 48.0 }, "checkpoint": "cc17c8cad21c52519bd34ad122ee6c1997e4a13099ad8967482fdb81a2692fa4", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s4e977af2af2584" ] }, { "age": 30, "six": 46.586666666666666, "five": 48.974, "tasks": { "blimp_filtered": 67.11, "supplement_filtered": 57.17, "ewok_filtered": 50.61, "entity_tracking": 18.57, "comps": 51.41, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 46.0 }, "checkpoint": "9637f31941f0ff3aed2cd0d17c5e2155c0ed367614396164638ec9a9257e8e61", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sd6b40e5056f52f" ] }, { "age": 40, "six": 46.700833333333335, "five": 49.008, "tasks": { "blimp_filtered": 68.0, "supplement_filtered": 56.46, "ewok_filtered": 51.19, "entity_tracking": 17.98, "comps": 51.41, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 48.0 }, "checkpoint": "67623618f45ab3af7831203e9c614e3e9bd95d6dac0cc2fe439feb4f57b365bc", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3e25a4d42833de" ] }, { "age": 50, "six": 46.95583333333333, "five": 49.126, "tasks": { "blimp_filtered": 68.62, "supplement_filtered": 55.74, "ewok_filtered": 51.17, "entity_tracking": 18.58, "comps": 51.52, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 46.0 }, "checkpoint": "379be6663dcf29dbb26917e8c5e98923b26474099ccd591b7a8f5222ad7ab75b", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6af7d5db37277e" ] }, { "age": 60, "six": 46.560833333333335, "five": 49.14, "tasks": { "blimp_filtered": 67.84, "supplement_filtered": 55.73, "ewok_filtered": 51.71, "entity_tracking": 18.56, "comps": 51.86, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 45.0 }, "checkpoint": "19ec97fd6d832b9a277b4cb50a1b0fde7a2e74b5296d7a70e0e3561c1078948e", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sf1c42daedf5a6d" ] }, { "age": 70, "six": 46.89, "five": 48.95, "tasks": { "blimp_filtered": 67.67, "supplement_filtered": 55.8, "ewok_filtered": 51.83, "entity_tracking": 17.97, "comps": 51.48, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 46.0 }, "checkpoint": "ba4542c4967c2a5d439f4ccd74926cd837f53c29c2747cb7be399666724bce6f", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sf03319de4ac5ea" ] }, { "age": 80, "six": 48.071666666666665, "five": 49.77, "tasks": { "blimp_filtered": 67.73, "supplement_filtered": 58.41, "ewok_filtered": 52.83, "entity_tracking": 18.56, "comps": 51.32, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 51.0 }, "checkpoint": "2b4f0555d6a5e614bd81d09555c008ac8f6310e75a6a3b96378e24fd9d5ff018", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5cafc5c72e7448" ] }, { "age": 90, "six": 46.665, "five": 48.968, "tasks": { "blimp_filtered": 67.34, "supplement_filtered": 55.9, "ewok_filtered": 51.99, "entity_tracking": 18.16, "comps": 51.45, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 47.0 }, "checkpoint": "b086ad3365932bf85555ba61941a81837b05a45469874d6a262df2c6fe8203d7", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "saf5aed7b9db3ee" ] }, { "age": 100, "six": 46.7025, "five": 49.016, "tasks": { "blimp_filtered": 67.41, "supplement_filtered": 56.27, "ewok_filtered": 51.89, "entity_tracking": 18.26, "comps": 51.25, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 46.0 }, "checkpoint": "d749187ce9da4fa809f95162b00dfcdfdaa64951734eb0552777af900bfc8501", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s46f8035aa67bbd" ] } ], "peak": { "five": { "score": 49.77, "ages": [ 80 ], "rank": 3 }, "six": { "score": 48.071666666666665, "ages": [ 80 ], "rank": 2 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 191, "id": "legacy-dense/base7", "original_name": "旧Dense・7層・分散なし", "title": "旧Dense・7層・分散なし", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 30, "six": 46.01, "five": 49.264, "tasks": { "blimp_filtered": 67.05, "supplement_filtered": 57.55, "ewok_filtered": 51.92, "entity_tracking": 18.31, "comps": 51.49, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 42.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s19a5b3447b53ae" ] }, { "age": 60, "six": 46.833333333333336, "five": 48.87, "tasks": { "blimp_filtered": 67.51, "supplement_filtered": 54.71, "ewok_filtered": 52.0, "entity_tracking": 18.38, "comps": 51.75, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s19a5b3447b53ae" ] }, { "age": 70, "six": 46.943333333333335, "five": 48.508, "tasks": { "blimp_filtered": 67.23, "supplement_filtered": 55.51, "ewok_filtered": 50.71, "entity_tracking": 17.69, "comps": 51.4, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 53.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s75b42a58ab3c1c" ] }, { "age": 80, "six": 48.0475, "five": 49.444, "tasks": { "blimp_filtered": 67.22, "supplement_filtered": 57.94, "ewok_filtered": 51.91, "entity_tracking": 18.39, "comps": 51.76, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 53.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s75b42a58ab3c1c" ] }, { "age": 90, "six": 46.9725, "five": 48.946, "tasks": { "blimp_filtered": 67.84, "supplement_filtered": 55.25, "ewok_filtered": 52.01, "entity_tracking": 18.22, "comps": 51.41, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s75b42a58ab3c1c" ] }, { "age": 100, "six": 46.94916666666666, "five": 48.906, "tasks": { "blimp_filtered": 67.29, "supplement_filtered": 56.54, "ewok_filtered": 51.64, "entity_tracking": 17.54, "comps": 51.52, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s19a5b3447b53ae" ] } ], "peak": { "five": { "score": 49.444, "ages": [ 80 ], "rank": 5 }, "six": { "score": 48.0475, "ages": [ 80 ], "rank": 3 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 192, "id": "legacy-dense/loop", "original_name": "旧Dense・分散全幅＋任意ループ", "title": "旧Dense・分散全幅＋任意ループ", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 46.211666666666666, "five": 48.424, "tasks": { "blimp_filtered": 66.28, "supplement_filtered": 57.68, "ewok_filtered": 50.09, "entity_tracking": 17.87, "comps": 50.2, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 47.0 }, "checkpoint": "6c87d80eec756e09dc35e3b9303484bf16f6079ed28e231bfca32146e2835f77", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sd88d5cc5ca89d7" ] }, { "age": 20, "six": 46.54333333333334, "five": 49.01, "tasks": { "blimp_filtered": 67.98, "supplement_filtered": 56.99, "ewok_filtered": 50.76, "entity_tracking": 18.55, "comps": 50.77, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 49.0 }, "checkpoint": "3b5fe98dc1d07220cfad357c42a0c1e4df7cd3878d882dd15167edf6dd3bdef8", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s19329b840eccd6" ] }, { "age": 30, "six": 46.63666666666666, "five": 49.322, "tasks": { "blimp_filtered": 66.66, "supplement_filtered": 59.38, "ewok_filtered": 51.03, "entity_tracking": 18.47, "comps": 51.07, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 47.0 }, "checkpoint": "6219655c8a2d26f6259f23188d9ebf89301c615f783a057db0308e435aec14fc", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sf3276ccc987af8" ] }, { "age": 40, "six": 47.235, "five": 49.358, "tasks": { "blimp_filtered": 68.89, "supplement_filtered": 56.9, "ewok_filtered": 51.42, "entity_tracking": 18.21, "comps": 51.37, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 48.0 }, "checkpoint": "69fbd1913daf835ae20fabc1d653103b6c6b45c5a59e629abea908b32ce359ff", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s57e23c8d9b55dd" ] }, { "age": 50, "six": 47.24916666666667, "five": 49.372, "tasks": { "blimp_filtered": 68.74, "supplement_filtered": 57.32, "ewok_filtered": 50.59, "entity_tracking": 18.48, "comps": 51.73, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 49.0 }, "checkpoint": "5467751e394f9ea34190fc7f31f02f83038b00551f33528e3a30a9d885e0b241", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sa15c45da085a9d" ] }, { "age": 60, "six": 46.57833333333333, "five": 49.064, "tasks": { "blimp_filtered": 68.65, "supplement_filtered": 55.85, "ewok_filtered": 51.38, "entity_tracking": 17.85, "comps": 51.59, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 45.0 }, "checkpoint": "9c3c1aa6301352caf9327c8bf20e4fc586e9ebbd11db9e06d1ac7710be1dbddc", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s4ab4a403e8ec03" ] }, { "age": 70, "six": 47.285, "five": 48.918, "tasks": { "blimp_filtered": 68.22, "supplement_filtered": 56.37, "ewok_filtered": 50.65, "entity_tracking": 17.92, "comps": 51.43, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 53.0 }, "checkpoint": "eb51c48874786e80e4f47912fd0dd1e0d86baf2b7d61d2d8035395c9cb02451a", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sde070c6b3934a7" ] }, { "age": 80, "six": 47.76, "five": 49.594, "tasks": { "blimp_filtered": 68.9, "supplement_filtered": 57.79, "ewok_filtered": 50.88, "entity_tracking": 19.28, "comps": 51.12, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 50.0 }, "checkpoint": "cba5f33f6947190bc661668abb426c52c2be3407f0317d5af6ab18161752458f", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s21310908773635" ] }, { "age": 90, "six": 46.94, "five": 49.11, "tasks": { "blimp_filtered": 67.66, "supplement_filtered": 57.77, "ewok_filtered": 50.01, "entity_tracking": 18.32, "comps": 51.79, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 45.0 }, "checkpoint": "da3be14a06a9b0e540778f19821fdfbdb4aface423e18750555a73b6e55b119f", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sf445f599b2ae36" ] }, { "age": 100, "six": 47.095, "five": 48.696, "tasks": { "blimp_filtered": 67.52, "supplement_filtered": 56.15, "ewok_filtered": 50.14, "entity_tracking": 17.87, "comps": 51.8, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 51.0 }, "checkpoint": "f37860cd536b9f9ca832fbcacade20fa3cc260fedace8107f08000ee83b2a4aa", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s0589814a139378" ] } ], "peak": { "five": { "score": 49.594, "ages": [ 80 ], "rank": 4 }, "six": { "score": 47.76, "ages": [ 80 ], "rank": 4 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 193, "id": "legacy-dense/gqa8", "original_name": "旧Dense・8層・GQA・FFN Router", "title": "旧Dense・8層・GQA・FFN Router", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.79833333333333, "five": 48.216, "tasks": { "blimp_filtered": 65.28, "supplement_filtered": 58.43, "ewok_filtered": 49.91, "entity_tracking": 17.55, "comps": 49.91, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scfd539023e5e95" ] }, { "age": 20, "six": 46.62166666666667, "five": 49.004, "tasks": { "blimp_filtered": 66.58, "supplement_filtered": 57.41, "ewok_filtered": 51.16, "entity_tracking": 18.93, "comps": 50.94, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 50.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scfd539023e5e95" ] }, { "age": 30, "six": 45.8625, "five": 48.59, "tasks": { "blimp_filtered": 65.97, "supplement_filtered": 55.33, "ewok_filtered": 51.93, "entity_tracking": 18.63, "comps": 51.09, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scfd539023e5e95" ] }, { "age": 40, "six": 46.22666666666667, "five": 49.322, "tasks": { "blimp_filtered": 67.42, "supplement_filtered": 57.1, "ewok_filtered": 52.19, "entity_tracking": 18.77, "comps": 51.13, "global_piqa_parallel": 16.5, "global_piqa_nonparallel": 45.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scfd539023e5e95" ] }, { "age": 50, "six": 46.3325, "five": 48.86, "tasks": { "blimp_filtered": 66.3, "supplement_filtered": 57.05, "ewok_filtered": 51.85, "entity_tracking": 18.12, "comps": 50.98, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scfd539023e5e95" ] }, { "age": 60, "six": 45.7475, "five": 48.752, "tasks": { "blimp_filtered": 66.52, "supplement_filtered": 55.72, "ewok_filtered": 52.48, "entity_tracking": 17.67, "comps": 51.37, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 43.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scfd539023e5e95" ] }, { "age": 70, "six": 45.608333333333334, "five": 48.582, "tasks": { "blimp_filtered": 66.42, "supplement_filtered": 55.92, "ewok_filtered": 51.24, "entity_tracking": 17.96, "comps": 51.37, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 44.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scfd539023e5e95" ] }, { "age": 80, "six": 46.57, "five": 49.054, "tasks": { "blimp_filtered": 67.09, "supplement_filtered": 57.38, "ewok_filtered": 51.49, "entity_tracking": 17.96, "comps": 51.35, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 45.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scfd539023e5e95" ] }, { "age": 90, "six": 47.6175, "five": 49.308, "tasks": { "blimp_filtered": 65.69, "supplement_filtered": 58.82, "ewok_filtered": 52.14, "entity_tracking": 18.3, "comps": 51.59, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 56.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scfd539023e5e95" ] }, { "age": 100, "six": 46.23, "five": 48.952, "tasks": { "blimp_filtered": 66.41, "supplement_filtered": 57.4, "ewok_filtered": 51.5, "entity_tracking": 18.38, "comps": 51.07, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 40.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scfd539023e5e95" ] } ], "peak": { "five": { "score": 49.322, "ages": [ 40 ], "rank": 10 }, "six": { "score": 47.6175, "ages": [ 90 ], "rank": 5 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 194, "id": "legacy-dense/base8", "original_name": "旧Dense・8層・MHA＋出力ゲート", "title": "旧Dense・8層・MHA＋出力ゲート", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.806666666666665, "five": 48.032, "tasks": { "blimp_filtered": 65.3, "supplement_filtered": 56.82, "ewok_filtered": 50.31, "entity_tracking": 17.68, "comps": 50.05, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s02e37eb8106946" ] }, { "age": 20, "six": 46.1725, "five": 48.668, "tasks": { "blimp_filtered": 67.15, "supplement_filtered": 57.89, "ewok_filtered": 50.11, "entity_tracking": 17.52, "comps": 50.67, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s02e37eb8106946" ] }, { "age": 30, "six": 46.21, "five": 48.81, "tasks": { "blimp_filtered": 67.02, "supplement_filtered": 56.52, "ewok_filtered": 50.79, "entity_tracking": 18.14, "comps": 51.58, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s02e37eb8106946" ] }, { "age": 40, "six": 47.025, "five": 49.4, "tasks": { "blimp_filtered": 67.74, "supplement_filtered": 57.22, "ewok_filtered": 52.24, "entity_tracking": 18.48, "comps": 51.32, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s02e37eb8106946" ] }, { "age": 50, "six": 47.19916666666666, "five": 49.318, "tasks": { "blimp_filtered": 67.9, "supplement_filtered": 57.25, "ewok_filtered": 51.59, "entity_tracking": 18.41, "comps": 51.44, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s02e37eb8106946" ] }, { "age": 60, "six": 46.79416666666667, "five": 48.532, "tasks": { "blimp_filtered": 66.61, "supplement_filtered": 54.99, "ewok_filtered": 51.7, "entity_tracking": 17.89, "comps": 51.47, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 50.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s02e37eb8106946" ] }, { "age": 70, "six": 47.61083333333333, "five": 49.412, "tasks": { "blimp_filtered": 67.51, "supplement_filtered": 56.9, "ewok_filtered": 52.66, "entity_tracking": 18.37, "comps": 51.62, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 51.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s02e37eb8106946" ] }, { "age": 80, "six": 47.244166666666665, "five": 49.372, "tasks": { "blimp_filtered": 67.44, "supplement_filtered": 56.82, "ewok_filtered": 52.11, "entity_tracking": 19.23, "comps": 51.26, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s02e37eb8106946" ] }, { "age": 90, "six": 47.34916666666667, "five": 49.186, "tasks": { "blimp_filtered": 67.68, "supplement_filtered": 54.84, "ewok_filtered": 52.91, "entity_tracking": 19.08, "comps": 51.42, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 54.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s02e37eb8106946" ] }, { "age": 100, "six": 46.994166666666665, "five": 48.866, "tasks": { "blimp_filtered": 67.06, "supplement_filtered": 55.48, "ewok_filtered": 51.29, "entity_tracking": 18.83, "comps": 51.67, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 51.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s02e37eb8106946" ] } ], "peak": { "five": { "score": 49.412, "ages": [ 70 ], "rank": 6 }, "six": { "score": 47.61083333333333, "ages": [ 70 ], "rank": 6 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 195, "id": "legacy-dense/min1", "original_name": "旧Dense・分散全幅＋最低1回ループ", "title": "旧Dense・分散全幅＋最低1回ループ", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.70916666666667, "five": 48.318, "tasks": { "blimp_filtered": 65.31, "supplement_filtered": 57.93, "ewok_filtered": 50.54, "entity_tracking": 18.16, "comps": 49.65, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 43.0 }, "checkpoint": "30a165b7a23bc5e3c24bc8636f0706b28bd0759b052b1bb0ce13ab76a58de87f", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sf1cd26405b53e2" ] }, { "age": 20, "six": 46.4725, "five": 49.022, "tasks": { "blimp_filtered": 67.64, "supplement_filtered": 56.84, "ewok_filtered": 51.66, "entity_tracking": 17.93, "comps": 51.04, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 49.0 }, "checkpoint": "77a12879566666ed7ad44493b40b5bbc0adafd46a0e88aaaea9beceeb51a477d", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s19562693524979" ] }, { "age": 30, "six": 46.25083333333333, "five": 48.762, "tasks": { "blimp_filtered": 66.64, "supplement_filtered": 57.15, "ewok_filtered": 51.22, "entity_tracking": 17.88, "comps": 50.92, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "25f8fcb4e09220a2ea64e3226e6899b529892f33f81876e77494fad6775e2a7e", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sa00d5d0eaa226d" ] }, { "age": 40, "six": 47.556666666666665, "five": 49.35, "tasks": { "blimp_filtered": 68.06, "supplement_filtered": 56.7, "ewok_filtered": 52.44, "entity_tracking": 18.11, "comps": 51.44, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 50.0 }, "checkpoint": "b4d826ec377f5c39271d716e0f20aae1a2c7b471290026e79666e56140d37913", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sf7e231be2b7f29" ] }, { "age": 50, "six": 46.6375, "five": 48.926, "tasks": { "blimp_filtered": 68.32, "supplement_filtered": 55.57, "ewok_filtered": 51.18, "entity_tracking": 18.0, "comps": 51.56, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 50.0 }, "checkpoint": "f4901be869411bb607690f95681a7996ea9a37a20f03793d1d8f9f26f325fadc", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6950de01b3403d" ] }, { "age": 60, "six": 46.27, "five": 48.5, "tasks": { "blimp_filtered": 68.02, "supplement_filtered": 53.13, "ewok_filtered": 51.53, "entity_tracking": 18.51, "comps": 51.31, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 45.0 }, "checkpoint": "1482d376a63f2369556071f629712bfd6e010e20939f1f1f3e1ef17788fb7f93", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5b2899a3820ca8" ] }, { "age": 70, "six": 46.51583333333333, "five": 48.486, "tasks": { "blimp_filtered": 68.4, "supplement_filtered": 54.03, "ewok_filtered": 51.41, "entity_tracking": 17.5, "comps": 51.09, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 51.0 }, "checkpoint": "56db09cca3ffd0d4f1f24208b9a07bff5159096895fc5a98827f9ea24eb31ef4", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sebfc59d9d0e830" ] }, { "age": 80, "six": 47.25833333333333, "five": 48.994, "tasks": { "blimp_filtered": 67.5, "supplement_filtered": 56.22, "ewok_filtered": 51.11, "entity_tracking": 18.5, "comps": 51.64, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 49.0 }, "checkpoint": "586f7b3244ba4c8dba884b8136654d9695037a7d991f18e332f36c576d5c8eb0", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s442626c64b1049" ] }, { "age": 90, "six": 47.2725, "five": 48.7, "tasks": { "blimp_filtered": 67.6, "supplement_filtered": 54.57, "ewok_filtered": 51.13, "entity_tracking": 18.65, "comps": 51.55, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 56.0 }, "checkpoint": "18de1a0bef9f30d2db2527fcbab3c2afc48ac8195e21755ff5b76f54777c3f39", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb4be1d1f3fdc1d" ] }, { "age": 100, "six": 47.01, "five": 48.888, "tasks": { "blimp_filtered": 68.07, "supplement_filtered": 54.82, "ewok_filtered": 51.53, "entity_tracking": 18.64, "comps": 51.38, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 50.0 }, "checkpoint": "daf2dbdd72ede38bd3a05f52463520a33d66f81bb0fce8fd6c0ca8043747cccb", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s9decf08a9efa57" ] } ], "peak": { "five": { "score": 49.35, "ages": [ 40 ], "rank": 7 }, "six": { "score": 47.556666666666665, "ages": [ 40 ], "rank": 7 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 196, "id": "legacy-dense/gated", "original_name": "旧Dense・分散全幅＋状態ゲート", "title": "旧Dense・分散全幅＋状態ゲート", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.549166666666665, "five": 48.032, "tasks": { "blimp_filtered": 66.09, "supplement_filtered": 55.99, "ewok_filtered": 50.61, "entity_tracking": 17.98, "comps": 49.49, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 42.0 }, "checkpoint": "e3f291d4cdf8e9882b77c150904b567065ebbb37466fc7b2ce8534a2dbc5ff82", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5f084f3cfcf9cc" ] }, { "age": 20, "six": 46.615, "five": 48.902, "tasks": { "blimp_filtered": 67.9, "supplement_filtered": 56.64, "ewok_filtered": 50.95, "entity_tracking": 18.27, "comps": 50.75, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 49.0 }, "checkpoint": "50ffe8225eebdba95f4f79e72a24a4adee7f0e32dbc765d176625600cfa7dcb8", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s9809e0749963cf" ] }, { "age": 30, "six": 46.17583333333334, "five": 48.572, "tasks": { "blimp_filtered": 67.13, "supplement_filtered": 56.02, "ewok_filtered": 50.31, "entity_tracking": 18.08, "comps": 51.32, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 48.0 }, "checkpoint": "138d382b3df2ba9d9737ddb8b4afc952b2d4facb52f9a9aa50182e1e810fde9a", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s9d3477a6798ee1" ] }, { "age": 40, "six": 46.4675, "five": 48.722, "tasks": { "blimp_filtered": 69.05, "supplement_filtered": 55.46, "ewok_filtered": 49.69, "entity_tracking": 18.17, "comps": 51.24, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 50.0 }, "checkpoint": "85ec5262a27b93df66e4e0dc6b427c68c8dca969cf61e409f56a8536085e5549", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sc910a7827b5610" ] }, { "age": 50, "six": 46.10583333333334, "five": 48.488, "tasks": { "blimp_filtered": 68.91, "supplement_filtered": 53.17, "ewok_filtered": 51.33, "entity_tracking": 17.55, "comps": 51.48, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 48.0 }, "checkpoint": "2d3d212fb03192fa13ae6363315efc0c13459b89ba775a9d71f19b17801111dd", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s064c3331a6017c" ] }, { "age": 60, "six": 47.32416666666666, "five": 48.668, "tasks": { "blimp_filtered": 67.81, "supplement_filtered": 53.02, "ewok_filtered": 52.06, "entity_tracking": 18.85, "comps": 51.6, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 55.0 }, "checkpoint": "25fcd2e8ab0b2bfc4e5f92977e5fecb2baeaadca744be0a1c7d30f442751d1df", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s778f7438b8d154" ] }, { "age": 70, "six": 47.55, "five": 49.036, "tasks": { "blimp_filtered": 68.5, "supplement_filtered": 54.51, "ewok_filtered": 52.0, "entity_tracking": 18.57, "comps": 51.6, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 55.0 }, "checkpoint": "961c511305a517c2d81c03b05310aba2eebea3bf8cb3a8e216b971d2f4266c6e", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5c4cec05e1c7f2" ] }, { "age": 80, "six": 47.07333333333333, "five": 49.058, "tasks": { "blimp_filtered": 67.66, "supplement_filtered": 56.25, "ewok_filtered": 51.17, "entity_tracking": 18.82, "comps": 51.39, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "911abc8441040a4962e5968fbc69ca63b8d2adc94aa5d9802f6c80542dfc8cb0", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s2e3cf2308fea0b" ] }, { "age": 90, "six": 47.4675, "five": 48.74, "tasks": { "blimp_filtered": 68.02, "supplement_filtered": 54.84, "ewok_filtered": 51.8, "entity_tracking": 17.83, "comps": 51.21, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 56.0 }, "checkpoint": "59320323ad844e711a8dc9763d375b7798bab313698d5d6407d26790cdc349f7", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sf582f4b1f8536e" ] }, { "age": 100, "six": 46.696666666666665, "five": 48.406, "tasks": { "blimp_filtered": 67.95, "supplement_filtered": 53.58, "ewok_filtered": 51.32, "entity_tracking": 17.41, "comps": 51.77, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 53.0 }, "checkpoint": "e0ba59d2dfb16fe575563d30751e6584d9913a73ac1642cbeadf78f0c57c02ee", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sf2813971a9feca" ] } ], "peak": { "five": { "score": 49.058, "ages": [ 80 ], "rank": 21 }, "six": { "score": 47.55, "ages": [ 70 ], "rank": 8 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 197, "id": "legacy-dense/return_plain", "original_name": "旧Dense・旧戻り先ループ・黄金角なし", "title": "旧Dense・旧戻り先ループ・黄金角なし", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 30, "six": 46.54, "five": 48.412, "tasks": { "blimp_filtered": 67.16, "supplement_filtered": 55.95, "ewok_filtered": 50.45, "entity_tracking": 17.66, "comps": 50.84, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 53.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb29c64bbb7079e" ] }, { "age": 60, "six": 46.428333333333335, "five": 48.584, "tasks": { "blimp_filtered": 67.12, "supplement_filtered": 55.07, "ewok_filtered": 51.42, "entity_tracking": 17.99, "comps": 51.32, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb29c64bbb7079e" ] }, { "age": 70, "six": 46.98416666666667, "five": 48.736, "tasks": { "blimp_filtered": 67.62, "supplement_filtered": 56.4, "ewok_filtered": 50.65, "entity_tracking": 17.72, "comps": 51.29, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 58.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb29c64bbb7079e" ] }, { "age": 80, "six": 47.388333333333335, "five": 48.742, "tasks": { "blimp_filtered": 67.29, "supplement_filtered": 55.86, "ewok_filtered": 51.23, "entity_tracking": 18.24, "comps": 51.09, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 56.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb29c64bbb7079e" ] }, { "age": 90, "six": 47.5125, "five": 49.294, "tasks": { "blimp_filtered": 67.04, "supplement_filtered": 58.23, "ewok_filtered": 51.06, "entity_tracking": 18.67, "comps": 51.47, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 51.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb29c64bbb7079e" ] }, { "age": 100, "six": 47.270833333333336, "five": 49.292, "tasks": { "blimp_filtered": 68.17, "supplement_filtered": 58.06, "ewok_filtered": 50.58, "entity_tracking": 18.03, "comps": 51.62, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb29c64bbb7079e" ] } ], "peak": { "five": { "score": 49.294, "ages": [ 90 ], "rank": 11 }, "six": { "score": 47.5125, "ages": [ 90 ], "rank": 9 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 198, "id": "legacy-dense/relation", "original_name": "旧Dense・現在状態×分散FiLM", "title": "旧Dense・現在状態×分散FiLM", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.369166666666665, "five": 47.91, "tasks": { "blimp_filtered": 65.52, "supplement_filtered": 55.35, "ewok_filtered": 50.58, "entity_tracking": 17.78, "comps": 50.32, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 43.0 }, "checkpoint": "eade46f531fab62e8a1e54e185717f3e40cd3aa7eff45a88792f848e12742301", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s57914885a3909b" ] }, { "age": 20, "six": 46.3575, "five": 48.69, "tasks": { "blimp_filtered": 67.6, "supplement_filtered": 56.69, "ewok_filtered": 49.78, "entity_tracking": 18.49, "comps": 50.89, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 49.0 }, "checkpoint": "468453afc1adb8d06ca75035d8f6e8c9da436f586ba7b1764979800ec2c833c4", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s7a7b386f2bd1b2" ] }, { "age": 30, "six": 45.9625, "five": 48.81, "tasks": { "blimp_filtered": 66.92, "supplement_filtered": 55.66, "ewok_filtered": 51.72, "entity_tracking": 18.54, "comps": 51.21, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 45.0 }, "checkpoint": "5a63ac11fa9c43482d08f5b78219ea96b00418fc42d7a79474b9a3d2c2494d36", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sc18bfffdf3793d" ] }, { "age": 40, "six": 46.405, "five": 49.05, "tasks": { "blimp_filtered": 68.23, "supplement_filtered": 55.56, "ewok_filtered": 51.91, "entity_tracking": 18.35, "comps": 51.2, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 45.0 }, "checkpoint": "1f14195922cf310923d4b28fedc20c7f225cc93d97058c570cfef68a46d9b3c7", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sded5f34901d6dd" ] }, { "age": 50, "six": 46.93, "five": 48.986, "tasks": { "blimp_filtered": 68.69, "supplement_filtered": 55.75, "ewok_filtered": 50.76, "entity_tracking": 18.56, "comps": 51.17, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "4875d1bb3c34d84f2e1592be70bb8a9b1889a9a60880bcfac09c703c4d527c50", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "se0b78a08130dc7" ] }, { "age": 60, "six": 46.66583333333333, "five": 48.878, "tasks": { "blimp_filtered": 67.49, "supplement_filtered": 56.0, "ewok_filtered": 50.89, "entity_tracking": 18.41, "comps": 51.6, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 45.0 }, "checkpoint": "bf28c81b556e2b73c437e2452674cbd68df4f599e756e619531b6c4cf6fa32e8", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sfc8a39a0d366cb" ] }, { "age": 70, "six": 46.55166666666667, "five": 48.532, "tasks": { "blimp_filtered": 68.4, "supplement_filtered": 54.8, "ewok_filtered": 49.71, "entity_tracking": 18.25, "comps": 51.5, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "cbbe36524c9ac822fa72b381fe67c27ff13fef84cf2c480635cf4fe97d0cb18a", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "se5780e816606d8" ] }, { "age": 80, "six": 47.46333333333333, "five": 49.04, "tasks": { "blimp_filtered": 68.15, "supplement_filtered": 55.92, "ewok_filtered": 50.9, "entity_tracking": 18.78, "comps": 51.45, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 51.0 }, "checkpoint": "ff6cf72e8e4816e55f6c9e9519a1332c2defa2df9075c1d272f0720075e54d27", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s575d3d23acb652" ] }, { "age": 90, "six": 47.208333333333336, "five": 49.126, "tasks": { "blimp_filtered": 68.38, "supplement_filtered": 56.02, "ewok_filtered": 50.73, "entity_tracking": 18.87, "comps": 51.63, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 50.0 }, "checkpoint": "01d5dec72a27fa1dff9c57b20c3a03e4d447ee1a786d97cd8d6346bfb92d8d86", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sff7ce08feee2ca" ] }, { "age": 100, "six": 46.47666666666667, "five": 48.842, "tasks": { "blimp_filtered": 67.98, "supplement_filtered": 56.55, "ewok_filtered": 49.76, "entity_tracking": 18.6, "comps": 51.32, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 46.0 }, "checkpoint": "d6e41e47a2fe5f7ee502cc98184fca74024765043c1a496b83cb8e8d21f67e6e", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "se3964798ee855c" ] } ], "peak": { "five": { "score": 49.126, "ages": [ 90 ], "rank": 17 }, "six": { "score": 47.46333333333333, "ages": [ 80 ], "rank": 10 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 199, "id": "legacy-dense/direct_ads", "original_name": "旧Dense・分散全幅＋ADS補正", "title": "旧Dense・分散全幅＋ADS補正", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.1275, "five": 47.726, "tasks": { "blimp_filtered": 65.96, "supplement_filtered": 55.9, "ewok_filtered": 49.69, "entity_tracking": 17.77, "comps": 49.31, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 40.0 }, "checkpoint": "c99f21815e0ad64e8717883fb5920871603bd14358f886bd913a1d36f5791d10", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sa7bcedf5fb42c1" ] }, { "age": 20, "six": 46.1475, "five": 48.444, "tasks": { "blimp_filtered": 67.12, "supplement_filtered": 55.8, "ewok_filtered": 50.42, "entity_tracking": 17.68, "comps": 51.2, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "39c2047654517799dfc8bbf87d1d21701d3f3a519a92eb81128fe8777c4f9d1d", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3a61787fada036" ] }, { "age": 30, "six": 46.041666666666664, "five": 48.314, "tasks": { "blimp_filtered": 66.42, "supplement_filtered": 56.65, "ewok_filtered": 49.78, "entity_tracking": 17.84, "comps": 50.88, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 48.0 }, "checkpoint": "fc41c386dabac37157ef536135633f6c96191d52548d438053534df0d1d24f1e", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s545ecd9633ed03" ] }, { "age": 40, "six": 47.39833333333333, "five": 49.048, "tasks": { "blimp_filtered": 68.41, "supplement_filtered": 57.03, "ewok_filtered": 50.51, "entity_tracking": 18.44, "comps": 50.85, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 55.0 }, "checkpoint": "73342bfe415087d6ef578345e4e62ae36b73458909619d9f554ae46e9b9e112f", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb83270b94a7fa0" ] }, { "age": 50, "six": 46.656666666666666, "five": 49.252, "tasks": { "blimp_filtered": 68.54, "supplement_filtered": 56.47, "ewok_filtered": 51.34, "entity_tracking": 18.66, "comps": 51.25, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 46.0 }, "checkpoint": "8c286c88cafe58d698b1bc48497cfcf4f5e24e437947002c91898a490a693d6e", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s99f86bed994be5" ] }, { "age": 60, "six": 46.54416666666667, "five": 48.726, "tasks": { "blimp_filtered": 67.51, "supplement_filtered": 55.44, "ewok_filtered": 51.24, "entity_tracking": 18.02, "comps": 51.42, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 47.0 }, "checkpoint": "da9f7f25b1b6e656f45b129820e67cddf6785a4f6f3339bbe08e2895003dff14", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sda3f6ddbe001e6" ] }, { "age": 70, "six": 46.314166666666665, "five": 48.838, "tasks": { "blimp_filtered": 66.96, "supplement_filtered": 57.45, "ewok_filtered": 50.88, "entity_tracking": 17.33, "comps": 51.57, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "d947433195851bbafaf4864aa59ce1ff2d588d0f8fa66de1cabcc4d2b84781c0", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "se96a145d2c9bf6" ] }, { "age": 80, "six": 47.4375, "five": 49.292, "tasks": { "blimp_filtered": 68.09, "supplement_filtered": 56.07, "ewok_filtered": 51.88, "entity_tracking": 18.69, "comps": 51.73, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 54.0 }, "checkpoint": "328af13fd94a4734de00a843e0fd965e72c51883cafb6ca59a6534d8e9ae4f99", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s1c2cb6b8619924" ] }, { "age": 90, "six": 46.486666666666665, "five": 48.954, "tasks": { "blimp_filtered": 67.06, "supplement_filtered": 55.65, "ewok_filtered": 52.07, "entity_tracking": 18.28, "comps": 51.71, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 45.0 }, "checkpoint": "6bb7daa997910b6415f780a47a1c31c7fe57793665badcdb52857a70b0a989cb", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s36fcd1d7e4acc0" ] }, { "age": 100, "six": 46.715833333333336, "five": 49.026, "tasks": { "blimp_filtered": 67.65, "supplement_filtered": 56.57, "ewok_filtered": 50.73, "entity_tracking": 18.65, "comps": 51.53, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 48.0 }, "checkpoint": "987f87e6aea6c31922db1c061979ec76e051c2603173a62a4a55508c518fcde0", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "se42c8417b339d0" ] } ], "peak": { "five": { "score": 49.292, "ages": [ 80 ], "rank": 12 }, "six": { "score": 47.4375, "ages": [ 80 ], "rank": 11 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 200, "id": "legacy-dense/depth10", "original_name": "旧Dense・10層", "title": "旧Dense・10層", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.85166666666667, "five": 49.28, "tasks": { "blimp_filtered": 67.77, "supplement_filtered": 58.2, "ewok_filtered": 50.5, "entity_tracking": 18.93, "comps": 51.0, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 50.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb8d0a0a0faf0d8" ] }, { "age": 30, "six": 47.089166666666664, "five": 48.886, "tasks": { "blimp_filtered": 67.22, "supplement_filtered": 55.98, "ewok_filtered": 51.01, "entity_tracking": 18.61, "comps": 51.61, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 50.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s2dd80e184a00a7" ] }, { "age": 40, "six": 47.05833333333333, "five": 49.922, "tasks": { "blimp_filtered": 68.78, "supplement_filtered": 57.71, "ewok_filtered": 52.3, "entity_tracking": 19.42, "comps": 51.4, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb8d0a0a0faf0d8" ] }, { "age": 50, "six": 47.38333333333333, "five": 49.23, "tasks": { "blimp_filtered": 68.45, "supplement_filtered": 55.86, "ewok_filtered": 51.52, "entity_tracking": 18.8, "comps": 51.52, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 53.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb8d0a0a0faf0d8" ] }, { "age": 60, "six": 46.8075, "five": 49.342, "tasks": { "blimp_filtered": 67.65, "supplement_filtered": 56.74, "ewok_filtered": 51.93, "entity_tracking": 18.64, "comps": 51.75, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 44.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s2dd80e184a00a7" ] }, { "age": 70, "six": 47.3225, "five": 49.06, "tasks": { "blimp_filtered": 68.05, "supplement_filtered": 56.16, "ewok_filtered": 51.04, "entity_tracking": 18.73, "comps": 51.32, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 53.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb8d0a0a0faf0d8" ] }, { "age": 80, "six": 47.04416666666667, "five": 49.12, "tasks": { "blimp_filtered": 67.56, "supplement_filtered": 56.21, "ewok_filtered": 51.86, "entity_tracking": 18.72, "comps": 51.25, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 51.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb8d0a0a0faf0d8" ] }, { "age": 90, "six": 47.115, "five": 49.102, "tasks": { "blimp_filtered": 67.27, "supplement_filtered": 55.94, "ewok_filtered": 51.27, "entity_tracking": 19.45, "comps": 51.58, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 53.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb8d0a0a0faf0d8" ] }, { "age": 100, "six": 46.64, "five": 49.032, "tasks": { "blimp_filtered": 67.73, "supplement_filtered": 55.01, "ewok_filtered": 51.93, "entity_tracking": 19.39, "comps": 51.1, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s2dd80e184a00a7" ] } ], "peak": { "five": { "score": 49.922, "ages": [ 40 ], "rank": 1 }, "six": { "score": 47.38333333333333, "ages": [ 50 ], "rank": 12 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 201, "id": "legacy-dense/direct_relation", "original_name": "旧Dense・分散全幅＋現在状態×分散FiLM", "title": "旧Dense・分散全幅＋現在状態×分散FiLM", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.72666666666667, "five": 48.33, "tasks": { "blimp_filtered": 65.79, "supplement_filtered": 57.7, "ewok_filtered": 50.43, "entity_tracking": 18.04, "comps": 49.69, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 46.0 }, "checkpoint": "3c82d575d812ae4691c6dd3bb20201fb674ee2e04ceffd03a45e909ffaa04c83", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5248cc38c2fab3" ] }, { "age": 20, "six": 46.219166666666666, "five": 48.624, "tasks": { "blimp_filtered": 67.05, "supplement_filtered": 56.67, "ewok_filtered": 50.37, "entity_tracking": 18.12, "comps": 50.91, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 48.0 }, "checkpoint": "ac686eefae9e52fbc71f2c95c790cae8dc3cf4bbff9ae4adcdda70d42594b855", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sbe995dfc3d5cb2" ] }, { "age": 30, "six": 46.035833333333336, "five": 48.51, "tasks": { "blimp_filtered": 66.81, "supplement_filtered": 56.06, "ewok_filtered": 50.43, "entity_tracking": 18.07, "comps": 51.18, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 45.0 }, "checkpoint": "175723b0588bf74f40fb6b874c5d035d64d7030736add0cd5275747056abae49", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb31356b978c596" ] }, { "age": 40, "six": 46.79833333333333, "five": 49.134, "tasks": { "blimp_filtered": 67.92, "supplement_filtered": 57.13, "ewok_filtered": 50.82, "entity_tracking": 18.57, "comps": 51.23, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 45.0 }, "checkpoint": "be15d73b5dd65f25570378bd7703eb13ee5b624bddb77560ad81a7ceb53f9900", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s0c284d03e34d55" ] }, { "age": 50, "six": 46.81166666666667, "five": 48.844, "tasks": { "blimp_filtered": 68.29, "supplement_filtered": 55.09, "ewok_filtered": 50.82, "entity_tracking": 18.59, "comps": 51.43, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "e0d97812d060343827e3d5be28cf3f7be5e378c6dfa7d560174ae8b464799043", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s4ebd36fbdde5ed" ] }, { "age": 60, "six": 46.85666666666667, "five": 48.698, "tasks": { "blimp_filtered": 68.2, "supplement_filtered": 54.65, "ewok_filtered": 51.46, "entity_tracking": 17.78, "comps": 51.4, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 52.0 }, "checkpoint": "09ed77dc8680e5057bead48b538f593fd0830787c3a8b6fdebc2be2bb7bd701d", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s9fc01cdd1cbee5" ] }, { "age": 70, "six": 46.42, "five": 48.968, "tasks": { "blimp_filtered": 68.4, "supplement_filtered": 55.84, "ewok_filtered": 50.96, "entity_tracking": 18.36, "comps": 51.28, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 46.0 }, "checkpoint": "5060c1dbf3f78323a536e616ffd2a4999a87ee2b0c59c9c4ecbc4b7c27d25a5b", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s9dc45a5744499a" ] }, { "age": 80, "six": 46.67666666666667, "five": 49.076, "tasks": { "blimp_filtered": 68.06, "supplement_filtered": 55.47, "ewok_filtered": 51.11, "entity_tracking": 19.44, "comps": 51.3, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 48.0 }, "checkpoint": "e919dd8e1440e23c8566c3ba046793133fa6a3c4bb53c5253b6bc495e9dd0735", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sfdd23bbac6b96e" ] }, { "age": 90, "six": 47.315, "five": 48.954, "tasks": { "blimp_filtered": 67.77, "supplement_filtered": 55.99, "ewok_filtered": 51.07, "entity_tracking": 18.63, "comps": 51.31, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 53.0 }, "checkpoint": "49e7e3203fa250f227a0ad424b781c293a527bc1bc4abb562bd394759d189bbe", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "se5bd31d0254116" ] }, { "age": 100, "six": 46.1575, "five": 48.556, "tasks": { "blimp_filtered": 67.96, "supplement_filtered": 54.15, "ewok_filtered": 50.55, "entity_tracking": 18.91, "comps": 51.21, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "bd3f05207d53e8b678393d4cf1db72785018b67580f149805a10e2865325ab2b", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s48d3d245d4cc73" ] } ], "peak": { "five": { "score": 49.134, "ages": [ 40 ], "rank": 16 }, "six": { "score": 47.315, "ages": [ 90 ], "rank": 13 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 202, "id": "legacy-dense/heads4", "original_name": "旧Dense・4×64head・7層", "title": "旧Dense・4×64head・7層", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.77916666666667, "five": 48.196, "tasks": { "blimp_filtered": 65.35, "supplement_filtered": 57.58, "ewok_filtered": 50.64, "entity_tracking": 17.32, "comps": 50.09, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6881d52dddf7a8" ] }, { "age": 20, "six": 46.788333333333334, "five": 49.21, "tasks": { "blimp_filtered": 66.32, "supplement_filtered": 57.69, "ewok_filtered": 51.91, "entity_tracking": 19.33, "comps": 50.8, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6881d52dddf7a8" ] }, { "age": 30, "six": 45.75833333333333, "five": 48.374, "tasks": { "blimp_filtered": 65.48, "supplement_filtered": 55.88, "ewok_filtered": 51.41, "entity_tracking": 18.17, "comps": 50.93, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 44.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6881d52dddf7a8" ] }, { "age": 40, "six": 46.0825, "five": 48.666, "tasks": { "blimp_filtered": 67.44, "supplement_filtered": 56.87, "ewok_filtered": 49.53, "entity_tracking": 18.2, "comps": 51.29, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 44.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6881d52dddf7a8" ] }, { "age": 50, "six": 46.84166666666667, "five": 49.086, "tasks": { "blimp_filtered": 67.49, "supplement_filtered": 56.69, "ewok_filtered": 51.28, "entity_tracking": 18.54, "comps": 51.43, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6881d52dddf7a8" ] }, { "age": 60, "six": 46.979166666666664, "five": 48.454, "tasks": { "blimp_filtered": 66.8, "supplement_filtered": 54.88, "ewok_filtered": 50.6, "entity_tracking": 18.51, "comps": 51.48, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 53.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6881d52dddf7a8" ] }, { "age": 70, "six": 46.7475, "five": 48.776, "tasks": { "blimp_filtered": 66.73, "supplement_filtered": 56.47, "ewok_filtered": 51.04, "entity_tracking": 18.45, "comps": 51.19, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6881d52dddf7a8" ] }, { "age": 80, "six": 47.30416666666667, "five": 48.938, "tasks": { "blimp_filtered": 66.52, "supplement_filtered": 56.35, "ewok_filtered": 51.4, "entity_tracking": 18.9, "comps": 51.52, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 54.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6881d52dddf7a8" ] }, { "age": 90, "six": 46.89666666666667, "five": 48.446, "tasks": { "blimp_filtered": 66.69, "supplement_filtered": 54.51, "ewok_filtered": 50.6, "entity_tracking": 18.8, "comps": 51.63, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 55.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6881d52dddf7a8" ] }, { "age": 100, "six": 47.08, "five": 48.672, "tasks": { "blimp_filtered": 66.58, "supplement_filtered": 55.87, "ewok_filtered": 50.58, "entity_tracking": 18.53, "comps": 51.8, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 53.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6881d52dddf7a8" ] } ], "peak": { "five": { "score": 49.21, "ages": [ 20 ], "rank": 13 }, "six": { "score": 47.30416666666667, "ages": [ 80 ], "rank": 14 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 203, "id": "legacy-dense/variance", "original_name": "旧Dense・分散低次元FiLM", "title": "旧Dense・分散低次元FiLM", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.7625, "five": 47.776, "tasks": { "blimp_filtered": 64.83, "supplement_filtered": 56.65, "ewok_filtered": 49.92, "entity_tracking": 17.26, "comps": 50.22, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 51.0 }, "checkpoint": "7abd24397d34db2eb5ac658cb47af577a81083ec069a2e161e52916b87830fe1", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s89300284431ad0" ] }, { "age": 20, "six": 46.01916666666666, "five": 48.378, "tasks": { "blimp_filtered": 67.31, "supplement_filtered": 55.02, "ewok_filtered": 50.33, "entity_tracking": 18.32, "comps": 50.91, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 50.0 }, "checkpoint": "f265685c026c840b1f8d1185d00e94a08940bb21aecf5c8a8eb0db3998e0c385", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6483d8ff99e445" ] }, { "age": 30, "six": 46.205, "five": 48.71, "tasks": { "blimp_filtered": 67.85, "supplement_filtered": 54.73, "ewok_filtered": 51.22, "entity_tracking": 18.4, "comps": 51.35, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 46.0 }, "checkpoint": "a80a995c1df548053f41dc6023774f71f5779965a0f52cd98a0fd036d82b2445", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sbdc9b489e8dc5c" ] }, { "age": 40, "six": 46.515, "five": 48.694, "tasks": { "blimp_filtered": 67.45, "supplement_filtered": 54.71, "ewok_filtered": 51.34, "entity_tracking": 18.26, "comps": 51.71, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 46.0 }, "checkpoint": "753999e3581f08318ebf4e7288a76e7a9ec7db2bc1c9e40002f8d66ed6377717", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s090314af37b564" ] }, { "age": 50, "six": 47.236666666666665, "five": 48.968, "tasks": { "blimp_filtered": 66.91, "supplement_filtered": 56.3, "ewok_filtered": 51.24, "entity_tracking": 18.68, "comps": 51.71, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 49.0 }, "checkpoint": "8dd535cddebcbf45917b4b5a13762f6ac914677613c6205cc227f863dcac729c", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s6e583fc0935ecd" ] }, { "age": 60, "six": 47.275, "five": 48.912, "tasks": { "blimp_filtered": 66.83, "supplement_filtered": 56.32, "ewok_filtered": 51.8, "entity_tracking": 18.25, "comps": 51.36, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 51.0 }, "checkpoint": "970612bfb3d9df585b581dbe2ed03804a656e8442e60afaa908c245a50ffef2c", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sf4307df26f1ed7" ] }, { "age": 70, "six": 46.766666666666666, "five": 49.09, "tasks": { "blimp_filtered": 67.52, "supplement_filtered": 57.38, "ewok_filtered": 51.39, "entity_tracking": 17.67, "comps": 51.49, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 47.0 }, "checkpoint": "e4071f992092d7d09ccde4e87b17053e5520739154f9f002ab41357057da6869", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s8094bc850981a7" ] }, { "age": 80, "six": 46.545833333333334, "five": 49.116, "tasks": { "blimp_filtered": 67.69, "supplement_filtered": 57.36, "ewok_filtered": 50.98, "entity_tracking": 18.12, "comps": 51.43, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "2c3b68c5217c4c0a5e1f21bc61b2b88d388d203d1cb3a2cacdd71359a0e85ba7", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s7588b25901f871" ] }, { "age": 90, "six": 46.916666666666664, "five": 49.07, "tasks": { "blimp_filtered": 67.71, "supplement_filtered": 56.54, "ewok_filtered": 51.15, "entity_tracking": 18.34, "comps": 51.61, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 49.0 }, "checkpoint": "89b064aa92c7b19f454cc2270e47c0e86bfac1c0ac01c8e7251041f749e1be01", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "saf0363791b3bc6" ] }, { "age": 100, "six": 46.5175, "five": 48.594, "tasks": { "blimp_filtered": 67.32, "supplement_filtered": 56.11, "ewok_filtered": 50.29, "entity_tracking": 17.63, "comps": 51.62, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 48.0 }, "checkpoint": "5836dbcc236e2d41499f5fa54b3577a57e1f985eda9c4e4f528f6ba3c9f07e96", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sf362550da6677f" ] } ], "peak": { "five": { "score": 49.116, "ages": [ 80 ], "rank": 18 }, "six": { "score": 47.275, "ages": [ 60 ], "rank": 15 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 204, "id": "legacy-dense/return_golden", "original_name": "旧Dense・旧戻り先ループ・黄金角あり", "title": "旧Dense・旧戻り先ループ・黄金角あり", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 30, "six": 46.42166666666667, "five": 48.67, "tasks": { "blimp_filtered": 65.75, "supplement_filtered": 57.2, "ewok_filtered": 50.66, "entity_tracking": 18.43, "comps": 51.31, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 49.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sd41671a85ec107" ] }, { "age": 60, "six": 46.74583333333333, "five": 49.068, "tasks": { "blimp_filtered": 68.68, "supplement_filtered": 54.61, "ewok_filtered": 52.95, "entity_tracking": 17.68, "comps": 51.42, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sd41671a85ec107" ] }, { "age": 70, "six": 47.20583333333333, "five": 48.62, "tasks": { "blimp_filtered": 68.23, "supplement_filtered": 54.27, "ewok_filtered": 52.13, "entity_tracking": 17.44, "comps": 51.03, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 56.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sd41671a85ec107" ] }, { "age": 80, "six": 47.06333333333333, "five": 48.96, "tasks": { "blimp_filtered": 66.98, "supplement_filtered": 56.89, "ewok_filtered": 50.69, "entity_tracking": 18.88, "comps": 51.36, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sd41671a85ec107" ] }, { "age": 90, "six": 47.001666666666665, "five": 48.872, "tasks": { "blimp_filtered": 67.38, "supplement_filtered": 54.66, "ewok_filtered": 52.67, "entity_tracking": 18.34, "comps": 51.31, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 52.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sd41671a85ec107" ] }, { "age": 100, "six": 46.653333333333336, "five": 49.342, "tasks": { "blimp_filtered": 67.8, "supplement_filtered": 56.65, "ewok_filtered": 52.67, "entity_tracking": 18.13, "comps": 51.46, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sd41671a85ec107" ] } ], "peak": { "five": { "score": 49.342, "ages": [ 100 ], "rank": 8 }, "six": { "score": 47.20583333333333, "ages": [ 70 ], "rank": 16 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 205, "id": "legacy-dense/width320", "original_name": "旧Dense・本体320・7層", "title": "旧Dense・本体320・7層", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.288333333333334, "five": 47.81, "tasks": { "blimp_filtered": 65.95, "supplement_filtered": 55.76, "ewok_filtered": 49.62, "entity_tracking": 17.7, "comps": 50.02, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 44.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scddb8827e68f2c" ] }, { "age": 20, "six": 46.778333333333336, "five": 49.092, "tasks": { "blimp_filtered": 67.46, "supplement_filtered": 56.82, "ewok_filtered": 51.38, "entity_tracking": 18.57, "comps": 51.23, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 51.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scddb8827e68f2c" ] }, { "age": 30, "six": 46.19583333333333, "five": 48.696, "tasks": { "blimp_filtered": 65.94, "supplement_filtered": 55.66, "ewok_filtered": 51.38, "entity_tracking": 18.77, "comps": 51.73, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scddb8827e68f2c" ] }, { "age": 40, "six": 47.031666666666666, "five": 49.208, "tasks": { "blimp_filtered": 68.37, "supplement_filtered": 56.03, "ewok_filtered": 51.89, "entity_tracking": 18.46, "comps": 51.29, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 49.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scddb8827e68f2c" ] }, { "age": 50, "six": 46.589166666666664, "five": 49.074, "tasks": { "blimp_filtered": 68.41, "supplement_filtered": 56.4, "ewok_filtered": 50.33, "entity_tracking": 18.39, "comps": 51.84, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scddb8827e68f2c" ] }, { "age": 60, "six": 45.76166666666666, "five": 48.178, "tasks": { "blimp_filtered": 67.28, "supplement_filtered": 53.03, "ewok_filtered": 50.96, "entity_tracking": 18.11, "comps": 51.51, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scddb8827e68f2c" ] }, { "age": 70, "six": 46.935, "five": 48.592, "tasks": { "blimp_filtered": 67.34, "supplement_filtered": 55.12, "ewok_filtered": 50.86, "entity_tracking": 17.91, "comps": 51.73, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 54.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scddb8827e68f2c" ] }, { "age": 80, "six": 46.59916666666667, "five": 48.986, "tasks": { "blimp_filtered": 67.42, "supplement_filtered": 56.87, "ewok_filtered": 51.1, "entity_tracking": 18.13, "comps": 51.41, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scddb8827e68f2c" ] }, { "age": 90, "six": 46.428333333333335, "five": 48.978, "tasks": { "blimp_filtered": 66.9, "supplement_filtered": 57.07, "ewok_filtered": 51.18, "entity_tracking": 18.12, "comps": 51.62, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scddb8827e68f2c" ] }, { "age": 100, "six": 46.56666666666667, "five": 48.95, "tasks": { "blimp_filtered": 66.94, "supplement_filtered": 56.45, "ewok_filtered": 50.89, "entity_tracking": 18.7, "comps": 51.77, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scddb8827e68f2c" ] } ], "peak": { "five": { "score": 49.208, "ages": [ 40 ], "rank": 14 }, "six": { "score": 47.031666666666666, "ages": [ 40 ], "rank": 17 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 206, "id": "legacy-dense/ads", "original_name": "旧Dense・ADS補正", "title": "旧Dense・ADS補正", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 46.016666666666666, "five": 48.09, "tasks": { "blimp_filtered": 65.07, "supplement_filtered": 57.24, "ewok_filtered": 50.42, "entity_tracking": 17.32, "comps": 50.4, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 48.0 }, "checkpoint": "fba3f6adce47cce65cd67d36511881ed29a53fce7cd0cb2a7f3e98e7bdad5528", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s4368d06581e83c" ] }, { "age": 20, "six": 46.681666666666665, "five": 48.882, "tasks": { "blimp_filtered": 66.79, "supplement_filtered": 58.04, "ewok_filtered": 50.1, "entity_tracking": 18.27, "comps": 51.21, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 50.0 }, "checkpoint": "172af51f4a14680685cba9b90f5fd701a2b8a4c39cc200c2f3f1e4dbefe9b604", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3f18cd9381f5bf" ] }, { "age": 30, "six": 46.26916666666666, "five": 49.084, "tasks": { "blimp_filtered": 67.21, "supplement_filtered": 56.77, "ewok_filtered": 51.52, "entity_tracking": 18.52, "comps": 51.4, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 44.0 }, "checkpoint": "bed3a34bc962534256d692c4f4f4e04d5a91ca63db5ad3eb2adcfedbf5967556", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "seca38185fa83d5" ] }, { "age": 40, "six": 46.715833333333336, "five": 48.932, "tasks": { "blimp_filtered": 68.87, "supplement_filtered": 55.74, "ewok_filtered": 50.62, "entity_tracking": 18.17, "comps": 51.26, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 47.0 }, "checkpoint": "1a77f35d0ddb3573bebe2d4688e9894fbb59833bec38da7fd7242c4fcd165340", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s44a4734486326f" ] }, { "age": 50, "six": 46.725833333333334, "five": 48.944, "tasks": { "blimp_filtered": 68.68, "supplement_filtered": 55.39, "ewok_filtered": 50.79, "entity_tracking": 18.57, "comps": 51.29, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 47.0 }, "checkpoint": "dc81eb9a943fa69e2e79746e9d55b408e6285cf25a0f4048fe0072203ee7037c", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sf4378d2bdae409" ] }, { "age": 60, "six": 46.37833333333333, "five": 48.618, "tasks": { "blimp_filtered": 68.08, "supplement_filtered": 54.55, "ewok_filtered": 50.67, "entity_tracking": 18.53, "comps": 51.26, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 49.0 }, "checkpoint": "2cda5fb25b7fe69fa5399fb9e66dc543265d81b96d91f9b4926277327e719fd8", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s27ba5c66622b99" ] }, { "age": 70, "six": 46.605, "five": 48.784, "tasks": { "blimp_filtered": 67.41, "supplement_filtered": 54.86, "ewok_filtered": 50.74, "entity_tracking": 19.0, "comps": 51.91, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 52.0 }, "checkpoint": "bb3e8dd6657745deff8786f6b8a5d5fb1519079a8447ba45fbe144df857958bf", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb3762e95806ea8" ] }, { "age": 80, "six": 46.9675, "five": 49.022, "tasks": { "blimp_filtered": 67.76, "supplement_filtered": 55.83, "ewok_filtered": 50.61, "entity_tracking": 19.03, "comps": 51.88, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 53.0 }, "checkpoint": "1ebc0942f8f0908e425ed9692b5251d5491b9c81bbe77231d2c9686e7200badc", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb59aec83f38c42" ] }, { "age": 90, "six": 46.65083333333333, "five": 48.16, "tasks": { "blimp_filtered": 67.17, "supplement_filtered": 53.0, "ewok_filtered": 50.56, "entity_tracking": 18.8, "comps": 51.27, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 52.0 }, "checkpoint": "af7e570234b2b53c8cca43ce97b7050914f9fb27787462aa69b88f45ceb500c1", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s409cf835ad5f60" ] }, { "age": 100, "six": 46.87916666666667, "five": 48.722, "tasks": { "blimp_filtered": 67.53, "supplement_filtered": 55.46, "ewok_filtered": 50.62, "entity_tracking": 18.43, "comps": 51.57, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 53.0 }, "checkpoint": "55f0d8cbfe1a38b1face5af6acb632120f60d83b30832e9b9ba5f6cf087cef39", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sef464e43e896aa" ] } ], "peak": { "five": { "score": 49.084, "ages": [ 30 ], "rank": 19 }, "six": { "score": 46.9675, "ages": [ 80 ], "rank": 18 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 207, "id": "legacy-dense/depth6", "original_name": "旧Dense・6層", "title": "旧Dense・6層", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 30, "six": 46.945, "five": 48.91, "tasks": { "blimp_filtered": 66.27, "supplement_filtered": 57.34, "ewok_filtered": 50.99, "entity_tracking": 18.54, "comps": 51.41, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 49.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sc3d53fa352dfc1" ] }, { "age": 60, "six": 46.08416666666667, "five": 48.574, "tasks": { "blimp_filtered": 66.49, "supplement_filtered": 56.52, "ewok_filtered": 50.82, "entity_tracking": 17.58, "comps": 51.46, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 43.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sc3d53fa352dfc1" ] }, { "age": 70, "six": 46.3475, "five": 48.684, "tasks": { "blimp_filtered": 67.18, "supplement_filtered": 56.55, "ewok_filtered": 50.91, "entity_tracking": 17.45, "comps": 51.33, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sc3d53fa352dfc1" ] }, { "age": 80, "six": 46.74583333333333, "five": 49.062, "tasks": { "blimp_filtered": 66.99, "supplement_filtered": 57.91, "ewok_filtered": 51.01, "entity_tracking": 17.74, "comps": 51.66, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sc3d53fa352dfc1" ] }, { "age": 90, "six": 46.1225, "five": 48.902, "tasks": { "blimp_filtered": 66.95, "supplement_filtered": 56.64, "ewok_filtered": 51.37, "entity_tracking": 17.98, "comps": 51.57, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sc3d53fa352dfc1" ] }, { "age": 100, "six": 46.704166666666666, "five": 48.712, "tasks": { "blimp_filtered": 66.58, "supplement_filtered": 55.31, "ewok_filtered": 51.57, "entity_tracking": 18.57, "comps": 51.53, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 51.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sc3d53fa352dfc1" ] } ], "peak": { "five": { "score": 49.062, "ages": [ 80 ], "rank": 20 }, "six": { "score": 46.945, "ages": [ 30 ], "rank": 19 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 208, "id": "legacy-dense/ffn2_5", "original_name": "旧Dense・FFN2.5倍・8層", "title": "旧Dense・FFN2.5倍・8層", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 30, "six": 46.30583333333333, "five": 48.828, "tasks": { "blimp_filtered": 67.27, "supplement_filtered": 56.59, "ewok_filtered": 50.74, "entity_tracking": 18.41, "comps": 51.13, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s85140f196219f9" ] }, { "age": 60, "six": 46.844166666666666, "five": 48.692, "tasks": { "blimp_filtered": 67.63, "supplement_filtered": 54.35, "ewok_filtered": 51.84, "entity_tracking": 18.52, "comps": 51.12, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 49.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s85140f196219f9" ] }, { "age": 100, "six": 46.01583333333333, "five": 48.58, "tasks": { "blimp_filtered": 66.62, "supplement_filtered": 55.65, "ewok_filtered": 51.4, "entity_tracking": 18.46, "comps": 50.77, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s85140f196219f9" ] } ], "peak": { "five": { "score": 48.828, "ages": [ 30 ], "rank": 24 }, "six": { "score": 46.844166666666666, "ages": [ 60 ], "rank": 20 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 209, "id": "legacy-dense/attn320", "original_name": "旧Dense・Attention320・7層", "title": "旧Dense・Attention320・7層", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 30, "six": 46.225, "five": 48.634, "tasks": { "blimp_filtered": 67.2, "supplement_filtered": 56.72, "ewok_filtered": 49.52, "entity_tracking": 18.14, "comps": 51.59, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s16673135b71b7c" ] }, { "age": 60, "six": 45.76583333333333, "five": 48.38, "tasks": { "blimp_filtered": 67.91, "supplement_filtered": 53.41, "ewok_filtered": 50.98, "entity_tracking": 17.84, "comps": 51.76, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 45.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s16673135b71b7c" ] }, { "age": 70, "six": 46.5175, "five": 49.082, "tasks": { "blimp_filtered": 67.72, "supplement_filtered": 56.65, "ewok_filtered": 51.28, "entity_tracking": 18.29, "comps": 51.47, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s16673135b71b7c" ] }, { "age": 80, "six": 46.755, "five": 49.17, "tasks": { "blimp_filtered": 67.51, "supplement_filtered": 56.36, "ewok_filtered": 51.33, "entity_tracking": 19.07, "comps": 51.58, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s16673135b71b7c" ] }, { "age": 90, "six": 46.295833333333334, "five": 48.922, "tasks": { "blimp_filtered": 68.26, "supplement_filtered": 56.04, "ewok_filtered": 50.9, "entity_tracking": 18.06, "comps": 51.35, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 44.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s16673135b71b7c" ] }, { "age": 100, "six": 46.30083333333334, "five": 48.522, "tasks": { "blimp_filtered": 67.35, "supplement_filtered": 53.67, "ewok_filtered": 51.26, "entity_tracking": 18.93, "comps": 51.4, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 50.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s16673135b71b7c" ] } ], "peak": { "five": { "score": 49.17, "ages": [ 80 ], "rank": 15 }, "six": { "score": 46.755, "ages": [ 80 ], "rank": 21 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 210, "id": "legacy-dense/fixed", "original_name": "旧Dense・分散全幅＋固定4〜7再実行", "title": "旧Dense・分散全幅＋固定4〜7再実行", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 44.82666666666667, "five": 47.45, "tasks": { "blimp_filtered": 66.27, "supplement_filtered": 54.8, "ewok_filtered": 48.81, "entity_tracking": 17.29, "comps": 50.08, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 44.0 }, "checkpoint": "a2dcec611d3aa82a1004fcbc9ebff13d645c7e66cb1298059e4b538525aad6a5", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s11f08d961607ad" ] }, { "age": 20, "six": 46.6725, "five": 48.774, "tasks": { "blimp_filtered": 68.15, "supplement_filtered": 55.77, "ewok_filtered": 51.09, "entity_tracking": 18.44, "comps": 50.42, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 50.0 }, "checkpoint": "1a22a85c7dc274f7961d772e4c9f9241eecfbbec0d88b4d0779bd358777c7fbe", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sb0eb0f5a4e6025" ] }, { "age": 30, "six": 45.4075, "five": 48.744, "tasks": { "blimp_filtered": 66.28, "supplement_filtered": 56.9, "ewok_filtered": 51.08, "entity_tracking": 18.25, "comps": 51.21, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 39.0 }, "checkpoint": "f0765a3c5830208da0986972152649300e5f2c2b58b540bfea0efdde831e6b0e", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "scfe1461484bd19" ] }, { "age": 40, "six": 46.48583333333333, "five": 49.338, "tasks": { "blimp_filtered": 68.75, "supplement_filtered": 57.31, "ewok_filtered": 51.25, "entity_tracking": 18.29, "comps": 51.09, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 46.0 }, "checkpoint": "278df0824f8cd703619accb5df14713b4f7d75dbcfd12b9d715f23080bad0a06", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sa97be641a3f1a6" ] }, { "age": 50, "six": 46.10583333333334, "five": 48.788, "tasks": { "blimp_filtered": 67.52, "supplement_filtered": 57.27, "ewok_filtered": 50.56, "entity_tracking": 17.48, "comps": 51.11, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 45.0 }, "checkpoint": "f0770f9a9ee7707a6c34d9ee122a96392660005698cee6c4ecdb93403d88658c", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "se1c94833ebd8d7" ] }, { "age": 60, "six": 46.020833333333336, "five": 48.586, "tasks": { "blimp_filtered": 67.97, "supplement_filtered": 54.29, "ewok_filtered": 51.27, "entity_tracking": 17.99, "comps": 51.41, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "bae665f1809346b3da11e9181be84f7e668198110deb2e4a39834d608740f959", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s40a7a4eb5cd83b" ] }, { "age": 70, "six": 46.31583333333333, "five": 48.746, "tasks": { "blimp_filtered": 68.11, "supplement_filtered": 55.4, "ewok_filtered": 51.11, "entity_tracking": 17.85, "comps": 51.26, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "87c35f3d2a3ca7f7c943691553a6d04efe5e3fdda0bfdb0a2e3d6017dfcb4668", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sfcba4dddbdfdac" ] }, { "age": 80, "six": 46.03666666666667, "five": 48.802, "tasks": { "blimp_filtered": 67.88, "supplement_filtered": 55.27, "ewok_filtered": 50.74, "entity_tracking": 18.87, "comps": 51.25, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 45.0 }, "checkpoint": "280fa57e84c48bdbe0da08c5f52281c571764a5d698bdc76ef2d9b6acbccd249", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s949959a05a4a8c" ] }, { "age": 90, "six": 46.181666666666665, "five": 48.882, "tasks": { "blimp_filtered": 67.68, "supplement_filtered": 55.95, "ewok_filtered": 51.14, "entity_tracking": 18.26, "comps": 51.38, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 44.0 }, "checkpoint": "0c4596deb1c9d96f7965e8996dbc98a80ca2905152b7d3cf31173722b4447422", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s0f5afcbe663dcd" ] }, { "age": 100, "six": 46.35666666666667, "five": 48.598, "tasks": { "blimp_filtered": 67.82, "supplement_filtered": 54.58, "ewok_filtered": 50.72, "entity_tracking": 18.37, "comps": 51.5, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 47.0 }, "checkpoint": "130b8f9e2d04086b0b8788d83c068d812358fd392903e0c5ef70c9640a0ac0bd", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s1606ebd757439a" ] } ], "peak": { "five": { "score": 49.338, "ages": [ 40 ], "rank": 9 }, "six": { "score": 46.6725, "ages": [ 20 ], "rank": 22 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 211, "id": "legacy-dense/ffn1_5", "original_name": "旧Dense・FFN1.5倍・8層", "title": "旧Dense・FFN1.5倍・8層", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 30, "six": 46.16166666666667, "five": 48.752, "tasks": { "blimp_filtered": 66.68, "supplement_filtered": 55.68, "ewok_filtered": 52.04, "entity_tracking": 18.04, "comps": 51.32, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s28707032b1154f" ] }, { "age": 60, "six": 46.465833333333336, "five": 48.52, "tasks": { "blimp_filtered": 68.28, "supplement_filtered": 54.55, "ewok_filtered": 49.56, "entity_tracking": 18.78, "comps": 51.43, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 52.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s28707032b1154f" ] }, { "age": 100, "six": 46.65833333333333, "five": 48.46, "tasks": { "blimp_filtered": 66.76, "supplement_filtered": 55.6, "ewok_filtered": 50.84, "entity_tracking": 17.96, "comps": 51.14, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 52.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s28707032b1154f" ] } ], "peak": { "five": { "score": 48.752, "ages": [ 30 ], "rank": 25 }, "six": { "score": 46.65833333333333, "ages": [ 100 ], "rank": 23 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 212, "id": "legacy-dense/ffn2", "original_name": "旧Dense・FFN2倍・8層", "title": "旧Dense・FFN2倍・8層", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 30, "six": 45.776666666666664, "five": 48.696, "tasks": { "blimp_filtered": 66.37, "supplement_filtered": 56.51, "ewok_filtered": 50.95, "entity_tracking": 18.17, "comps": 51.48, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 41.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s2195f6092b4c87" ] }, { "age": 60, "six": 46.376666666666665, "five": 48.528, "tasks": { "blimp_filtered": 66.98, "supplement_filtered": 54.66, "ewok_filtered": 50.92, "entity_tracking": 18.62, "comps": 51.46, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s2195f6092b4c87" ] }, { "age": 100, "six": 46.6125, "five": 49.008, "tasks": { "blimp_filtered": 67.85, "supplement_filtered": 55.34, "ewok_filtered": 51.02, "entity_tracking": 19.18, "comps": 51.65, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 45.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s2195f6092b4c87" ] } ], "peak": { "five": { "score": 49.008, "ages": [ 100 ], "rank": 22 }, "six": { "score": 46.6125, "ages": [ 100 ], "rank": 24 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 213, "id": "legacy-dense/depth12", "original_name": "旧Dense・12層", "title": "旧Dense・12層", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-dense", "cohort_label": "旧Baby・参考", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 30, "six": 46.365833333333335, "five": 49.0, "tasks": { "blimp_filtered": 67.48, "supplement_filtered": 56.75, "ewok_filtered": 50.55, "entity_tracking": 18.34, "comps": 51.88, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s04c0863e76502e" ] }, { "age": 60, "six": 46.10666666666667, "five": 48.698, "tasks": { "blimp_filtered": 68.92, "supplement_filtered": 54.61, "ewok_filtered": 50.3, "entity_tracking": 18.03, "comps": 51.63, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 43.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s04c0863e76502e" ] }, { "age": 100, "six": 45.76166666666666, "five": 48.572, "tasks": { "blimp_filtered": 67.87, "supplement_filtered": 53.68, "ewok_filtered": 52.02, "entity_tracking": 17.89, "comps": 51.4, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 44.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s04c0863e76502e" ] } ], "peak": { "five": { "score": 49.0, "ages": [ 30 ], "rank": 23 }, "six": { "score": 46.365833333333335, "ages": [ 30 ], "rank": 25 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 214, "id": "baby-boundary/B_bos_skip", "original_name": "B・最初から境界遮断", "title": "B・最初から境界遮断", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "baby-boundary", "cohort_label": "旧Baby・境界条件", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.15833333333333, "five": 47.848, "tasks": { "blimp_filtered": 63.92, "supplement_filtered": 58.06, "ewok_filtered": 49.86, "entity_tracking": 17.79, "comps": 49.61, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 44.0 }, "checkpoint": "70a84604491711c741c31b9f1bc4ca283ef0849d5bccf563eb770547ae13ee25", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s948ad4690e8a9a" ] }, { "age": 20, "six": 45.49166666666667, "five": 48.448, "tasks": { "blimp_filtered": 66.11, "supplement_filtered": 56.44, "ewok_filtered": 50.41, "entity_tracking": 18.54, "comps": 50.74, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 42.0 }, "checkpoint": "26424c6ab10333222becb581d52094b65e51aebdecbb2fbd3ca61e8b2f82f1a3", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sa6eea21259dac8" ] }, { "age": 30, "six": 45.985, "five": 48.552, "tasks": { "blimp_filtered": 65.88, "supplement_filtered": 55.21, "ewok_filtered": 51.7, "entity_tracking": 19.25, "comps": 50.72, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 43.0 }, "checkpoint": "1e846a1904bb5e46cad8c7815c105157d543bda10bbcd44ba6521bb40c5b0fa0", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s64231fc7d15faf" ] }, { "age": 40, "six": 46.045833333333334, "five": 48.322, "tasks": { "blimp_filtered": 65.79, "supplement_filtered": 56.16, "ewok_filtered": 50.96, "entity_tracking": 17.13, "comps": 51.57, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "d94ccc980e81d71275f0d76ca38d153d401802d52206400314e2b2cd113b52b8", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sd5437d81e40323" ] }, { "age": 50, "six": 46.0975, "five": 48.296, "tasks": { "blimp_filtered": 66.51, "supplement_filtered": 54.08, "ewok_filtered": 51.84, "entity_tracking": 17.77, "comps": 51.28, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 44.0 }, "checkpoint": "8eb0b6c3924a7e81486c7167449c780acb612a2e822597705daa2d761d854cba", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s38f87bcd4516fa" ] }, { "age": 60, "six": 45.089166666666664, "five": 47.468, "tasks": { "blimp_filtered": 65.23, "supplement_filtered": 52.68, "ewok_filtered": 49.96, "entity_tracking": 18.02, "comps": 51.45, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "d290201439f5195508dc3f5d1ad52044273f733234f1f3c702cf26c89c3d5b24", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s7fbb98d9daa9b6" ] }, { "age": 70, "six": 45.560833333333335, "five": 47.44, "tasks": { "blimp_filtered": 63.85, "supplement_filtered": 53.23, "ewok_filtered": 51.16, "entity_tracking": 17.63, "comps": 51.33, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 50.0 }, "checkpoint": "5c9d674b882fb183e11de3a253f410dcdffb4bb418feb05f5007d128e7fb3d11", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sce8fd9463d5404" ] }, { "age": 80, "six": 46.55, "five": 47.942, "tasks": { "blimp_filtered": 66.17, "supplement_filtered": 52.43, "ewok_filtered": 51.76, "entity_tracking": 18.01, "comps": 51.34, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 52.0 }, "checkpoint": "1fd0f258d1401e92519389cb7abc3b6a5104241cea9754667bf1406630dc4279", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sb8c08ef6ed64f3" ] }, { "age": 90, "six": 45.495, "five": 47.378, "tasks": { "blimp_filtered": 64.67, "supplement_filtered": 51.84, "ewok_filtered": 50.8, "entity_tracking": 18.07, "comps": 51.51, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 44.0 }, "checkpoint": "aa02ef76bceea4e0bbe2874ac59b9cc6fd19d6071d0522a37c8e6bd8d2c33ee5", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s9d48a7d1e41c74" ] }, { "age": 100, "six": 45.58416666666667, "five": 47.374, "tasks": { "blimp_filtered": 65.51, "supplement_filtered": 51.86, "ewok_filtered": 50.14, "entity_tracking": 17.98, "comps": 51.38, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 49.0 }, "checkpoint": "a62806cea4f8eba092e0458ab0a4bdb63fe8b03275a3b17ab3cb09658b48b2b2", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sd3cd0a5f6dba7d" ] } ], "peak": { "five": { "score": 48.552, "ages": [ 30 ], "rank": 1 }, "six": { "score": 46.55, "ages": [ 80 ], "rank": 1 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 215, "id": "baby-boundary/A_control", "original_name": "A・通常（元VISTARCと共通）", "title": "A・通常（元VISTARCと共通）", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "baby-boundary", "cohort_label": "旧Baby・境界条件", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.885, "five": 48.12, "tasks": { "blimp_filtered": 64.85, "supplement_filtered": 59.0, "ewok_filtered": 49.85, "entity_tracking": 17.29, "comps": 49.61, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 50.0 }, "checkpoint": "3128c6d579544bf969a2a036d10fbb7d501acc0a9de15081228f0854d053bcb3", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s8d1f550f0af4f2" ] }, { "age": 20, "six": 45.0325, "five": 47.786, "tasks": { "blimp_filtered": 65.39, "supplement_filtered": 55.35, "ewok_filtered": 49.14, "entity_tracking": 18.77, "comps": 50.28, "global_piqa_parallel": 15.53, "global_piqa_nonparallel": 47.0 }, "checkpoint": "b36c01dd3f00c655197d700791f4554f4e6747cc08522082df62c1859c9fa331", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s9f2be5aa66ded1" ] } ], "peak": { "five": { "score": 48.12, "ages": [ 10 ], "rank": 2 }, "six": { "score": 45.885, "ages": [ 10 ], "rank": 2 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 216, "id": "baby-boundary/E_delayed_bos_skip", "original_name": "E・10M以降境界遮断", "title": "E・10M以降境界遮断", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "baby-boundary", "cohort_label": "旧Baby・境界条件", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.885, "five": 48.12, "tasks": { "blimp_filtered": 64.85, "supplement_filtered": 59.0, "ewok_filtered": 49.85, "entity_tracking": 17.29, "comps": 49.61, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 50.0 }, "checkpoint": "5123d263d1f8a02cc5733199328da9c330bd87ed53cac18fcf64dd71d082f8ea", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sad390e0edd57bc" ] }, { "age": 20, "six": 45.58583333333333, "five": 47.958, "tasks": { "blimp_filtered": 65.48, "supplement_filtered": 56.78, "ewok_filtered": 48.27, "entity_tracking": 18.9, "comps": 50.36, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 49.0 }, "checkpoint": "28c5aee6233a8f2c5c9d6c25cc9eac345d00593a011cfb4b46ddefa4caf2c2c7", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "scd402e9a536801" ] }, { "age": 30, "six": 45.85, "five": 47.59, "tasks": { "blimp_filtered": 65.47, "supplement_filtered": 53.05, "ewok_filtered": 50.09, "entity_tracking": 18.68, "comps": 50.66, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "c9812a516f53659789ebe194d588239bbf351f6206c3d2b166923e0443f854d0", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "saca8f5ccb3859e" ] }, { "age": 40, "six": 45.026666666666664, "five": 47.196, "tasks": { "blimp_filtered": 64.25, "supplement_filtered": 53.01, "ewok_filtered": 49.96, "entity_tracking": 17.4, "comps": 51.36, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 47.0 }, "checkpoint": "f186f91e839a257af8c48970d67cea36bec2df83c62f24f3e54453c8a3ee0f2c", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s0617c2edc014eb" ] } ], "peak": { "five": { "score": 48.12, "ages": [ 10 ], "rank": 2 }, "six": { "score": 45.885, "ages": [ 10 ], "rank": 2 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 217, "id": "baby-boundary/D_bos_skip_repeat_half", "original_name": "D・境界遮断＋半減", "title": "D・境界遮断＋半減", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "baby-boundary", "cohort_label": "旧Baby・境界条件", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 45.345, "five": 47.966, "tasks": { "blimp_filtered": 64.23, "supplement_filtered": 56.08, "ewok_filtered": 50.41, "entity_tracking": 18.34, "comps": 50.77, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 47.0 }, "checkpoint": "0709fcaecae3af7089c8d7a56f586f94afd3197ad6eb5fd9c21dd48a72e0aab8", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s3ec901fcc1e766" ] } ], "peak": { "five": { "score": 47.966, "ages": [ 20 ], "rank": 4 }, "six": { "score": 45.345, "ages": [ 20 ], "rank": 4 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 218, "id": "baby-boundary/C_repeat_half", "original_name": "C・10M以降feedback半減", "title": "C・10M以降feedback半減", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "baby-boundary", "cohort_label": "旧Baby・境界条件", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 44.836666666666666, "five": 47.656, "tasks": { "blimp_filtered": 64.46, "supplement_filtered": 55.45, "ewok_filtered": 49.07, "entity_tracking": 18.66, "comps": 50.64, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 44.0 }, "checkpoint": "a0f917b5c9c7ecb88c27ec6b0da8ec7a6ea56f68b7011f327b8327f0c4f6357e", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s998f4d04710fdd" ] } ], "peak": { "five": { "score": 47.656, "ages": [ 20 ], "rank": 5 }, "six": { "score": 44.836666666666666, "ages": [ 20 ], "rank": 5 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 219, "id": "run/IPRLM320_FROZEN_PARENT_READOUT_20261012", "original_name": "IPRLM320_FROZEN_PARENT_READOUT_20261012", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby-frozen-body-continuation", "cohort_label": "Baby・本体凍結後の追加学習", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": { "input": 2253, "hidden": 320, "outputs": 3, "eps": 1e-06, "normalization": "seven RMS-normalized320 blocks + seven raw RMS + three entropy + three JS", "initialization": "local RNG fan_in up; zero down", "feedback": "original parent mixture only", "prediction": "softmax(parent_log_pi + residual)" }, "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "frozen_parent_mos3_residual_readout", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "readout_gate": { "input": 2253, "hidden": 320, "outputs": 3, "eps": 1e-06, "normalization": "seven RMS-normalized320 blocks + seven raw RMS + three entropy + three JS", "initialization": "local RNG fan_in up; zero down", "feedback": "original parent mixture only", "prediction": "softmax(parent_log_pi + residual)" }, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict prefix recurrence; frozen original internal feedback; output-only gate; original inputs and rankers", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true, "compile_readout_statistics": true }, "benchmark_words": [ 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 90000000, 100000000 ], "timing": "serial after100M", "parent80M_six_axis": 49.8675, "parent80M_five_axis": 52.002, "no_endpoint_penalty": true }, "milestones_words": [ 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": "mean/loss tiles and statistics dynamic; attention gate compiled; backbone uncompiled", "compile_dynamic": "mean/loss tiles and statistics dynamic", "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 82, "sources": [ "sd150f5eca477f0", "s47bf74cf9f4d04", "s8dba275be5d684" ], "points": [ { "age": 90.0, "six": 49.85916666666667, "five": 51.992, "tasks": { "blimp_filtered": 69.64, "supplement_filtered": 63.46, "ewok_filtered": 51.39, "entity_tracking": 23.16, "comps": 52.31, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 58.0 }, "checkpoint": "fce63fd4992a7277374e3a002263969e3c35c5237371cec6a1cf84699e7a8081", "precision": "fp32", "verification": "primary", "sources": [ "s21ed74f36c5b1e", "s74e7f44163ba5a", "sc23adb32b7d95c", "s4647dce6ddf0e0" ] }, { "age": 100.0, "six": 49.83416666666667, "five": 51.962, "tasks": { "blimp_filtered": 69.57, "supplement_filtered": 63.89, "ewok_filtered": 50.93, "entity_tracking": 23.08, "comps": 52.34, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 58.0 }, "checkpoint": "63bd68f8ba1f7e9ae63a25702f90df201943ac8750b87e2031be92ca64ceedc4", "precision": "fp32", "verification": "primary", "sources": [ "s95943b82332555", "s004905e5fe236b", "sc23adb32b7d95c", "s4647dce6ddf0e0" ] } ], "peak": { "five": { "score": 51.992, "ages": [ 90.0 ], "rank": 1 }, "six": { "score": 49.85916666666667, "ages": [ 90.0 ], "rank": 1 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 220, "id": "run/IPRLM320_FROZEN_SEGMENT_SELECTOR_20261012", "original_name": "IPRLM320_FROZEN_SEGMENT_SELECTOR_20261012", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "契約設定（詳細）", "cohort": "baby-frozen-body-continuation", "cohort_label": "Baby・本体凍結後の追加学習", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": { "input": 2253, "hidden": 320, "outputs": 3, "eps": 1e-06, "normalization": "seven RMS-normalized320 blocks + seven raw RMS + three entropy + three JS", "initialization": "local RNG fan_in up; zero down; initial hard tie selects branch1", "feedback": "original frozen parent mixture only", "prediction": "argmax of standalone selector scores; one component distribution for eight predictions; reset per document", "segment_tokens": 8, "teacher": "same actual continuation, detached per-branch mean NLL", "loss": "token-count and NLL-gap weighted all-pairs logistic preference; normalized by total gap mass" }, "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [], "state": "STOPPED_BY_USER", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "frozen_parent_mos3_segment_selector", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "readout_gate": { "input": 2253, "hidden": 320, "outputs": 3, "eps": 1e-06, "normalization": "seven RMS-normalized320 blocks + seven raw RMS + three entropy + three JS", "initialization": "local RNG fan_in up; zero down; initial hard tie selects branch1", "feedback": "original frozen parent mixture only", "prediction": "argmax of standalone selector scores; one component distribution for eight predictions; reset per document", "segment_tokens": 8, "teacher": "same actual continuation, detached per-branch mean NLL", "loss": "token-count and NLL-gap weighted all-pairs logistic preference; normalized by total gap mass" }, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 exact parent prefix recurrence; output hard one-branch segment selection; original questions and rankers", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": true, "compile_readout_statistics": true }, "benchmark_words": [ 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 90000000, 100000000 ], "timing": "serial after100M", "parent80M_six_axis": 49.8675, "parent80M_five_axis": 52.002, "no_endpoint_penalty": true }, "milestones_words": [ 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": "selector distribution statistics and attention received gate only; parent mean/loss tiles not compiled", "compile_dynamic": "selector distribution statistics", "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 83, "sources": [ "sc23aaf09b8efd6", "s1f7572e736e4d6", "s574bef6d95b6fb" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 221, "id": "baby-boundary/ablation_100M_feedback_zero", "original_name": "元100M・推論feedback OFF", "title": "元100M・推論feedback OFF", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "baby-inference-only", "cohort_label": "Baby・推論時介入", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 100, "six": 44.27333333333333, "five": 46.498, "tasks": { "blimp_filtered": 64.41, "supplement_filtered": 48.65, "ewok_filtered": 50.62, "entity_tracking": 17.92, "comps": 50.89, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 43.0 }, "checkpoint": "afac36e9892d31530cde8b33f913092d24d7da25181690d9d12a8c5462f41be3", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s6bdff8c7f4f1a8" ] } ], "peak": { "five": { "score": 46.498, "ages": [ 100 ], "rank": 2 }, "six": { "score": 44.27333333333333, "ages": [ 100 ], "rank": 1 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 222, "id": "baby-boundary/ablation_100M_bos_skip", "original_name": "元100M・推論境界遮断", "title": "元100M・推論境界遮断", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "baby-inference-only", "cohort_label": "Baby・推論時介入", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 100, "six": 44.23, "five": 46.546, "tasks": { "blimp_filtered": 64.37, "supplement_filtered": 48.58, "ewok_filtered": 51.05, "entity_tracking": 17.83, "comps": 50.9, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 42.0 }, "checkpoint": "afac36e9892d31530cde8b33f913092d24d7da25181690d9d12a8c5462f41be3", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sf043fe3cb2102e" ] } ], "peak": { "five": { "score": 46.546, "ages": [ 100 ], "rank": 1 }, "six": { "score": 44.23, "ages": [ 100 ], "rank": 2 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 223, "id": "baby-boundary/ablation_100M_feedback_half", "original_name": "元100M・推論feedback半減", "title": "元100M・推論feedback半減", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "baby-inference-only", "cohort_label": "Baby・推論時介入", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 100, "six": 44.18333333333333, "five": 46.49, "tasks": { "blimp_filtered": 64.38, "supplement_filtered": 48.38, "ewok_filtered": 50.96, "entity_tracking": 17.84, "comps": 50.89, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 42.0 }, "checkpoint": "afac36e9892d31530cde8b33f913092d24d7da25181690d9d12a8c5462f41be3", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1948008f116c34" ] } ], "peak": { "five": { "score": 46.49, "ages": [ 100 ], "rank": 3 }, "six": { "score": 44.18333333333333, "ages": [ 100 ], "rank": 3 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 224, "id": "baby-original/original_vistarc", "original_name": "元VISTARC", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "baby-original", "cohort_label": "旧Baby・元設定", "seed": null, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0 }, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 10, "six": 45.885, "five": 48.12, "tasks": { "blimp_filtered": 64.85, "supplement_filtered": 59.0, "ewok_filtered": 49.85, "entity_tracking": 17.29, "comps": 49.61, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 50.0 }, "checkpoint": "0ccae40ed3a42a70c747486027d787971b0753d2d2030e2cbb1fd531739dbe3f", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s07128033ac4a4c" ] }, { "age": 20, "six": 45.0325, "five": 47.786, "tasks": { "blimp_filtered": 65.39, "supplement_filtered": 55.35, "ewok_filtered": 49.14, "entity_tracking": 18.77, "comps": 50.28, "global_piqa_parallel": 15.53, "global_piqa_nonparallel": 47.0 }, "checkpoint": "be0f98c52940c4241bf4783734463bfad3bd8969387ce00c4511b68ad5539f12", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sa8c89cdc3091cd" ] }, { "age": 30, "six": 45.265, "five": 47.482, "tasks": { "blimp_filtered": 65.27, "supplement_filtered": 53.49, "ewok_filtered": 49.97, "entity_tracking": 18.08, "comps": 50.6, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 47.0 }, "checkpoint": "50f60a24d6e3830b588ac6d582a423831cec5ebd724b0e8cd0c8dc484bd261e4", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s1b3711e1961344" ] }, { "age": 40, "six": 45.305, "five": 47.33, "tasks": { "blimp_filtered": 64.47, "supplement_filtered": 52.65, "ewok_filtered": 50.58, "entity_tracking": 17.97, "comps": 50.98, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 49.0 }, "checkpoint": "e4491ae483e48840a059a1134f979a1bfe62787114375a54911bcae6e66abe28", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "se3267c317d6b1b" ] }, { "age": 50, "six": 45.598333333333336, "five": 47.288, "tasks": { "blimp_filtered": 63.96, "supplement_filtered": 52.98, "ewok_filtered": 50.8, "entity_tracking": 17.61, "comps": 51.09, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "f976d04bcf97c56bf568fe5c99a8ebebda12536e01d60e7eb2eccf8d4abad680", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sdf1ca5f7b99f4e" ] }, { "age": 60, "six": 45.513333333333335, "five": 47.092, "tasks": { "blimp_filtered": 64.92, "supplement_filtered": 51.75, "ewok_filtered": 50.58, "entity_tracking": 17.3, "comps": 50.91, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 50.0 }, "checkpoint": "4a1a8887701065eaa5cbfe7928b3ed8447b76a63baa43b851b9c822f219b5d4b", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sa8ddbab3f58b0f" ] }, { "age": 70, "six": 45.49583333333333, "five": 47.156, "tasks": { "blimp_filtered": 64.75, "supplement_filtered": 52.55, "ewok_filtered": 50.13, "entity_tracking": 17.46, "comps": 50.89, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 54.0 }, "checkpoint": "9fdca19ccddb4f8cdd9ec38f076a0318e7765357e71a873f11616b1ed9c9e3ac", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sabf492751b04ae" ] }, { "age": 80, "six": 45.18333333333333, "five": 47.284, "tasks": { "blimp_filtered": 64.39, "supplement_filtered": 52.05, "ewok_filtered": 51.31, "entity_tracking": 17.71, "comps": 50.96, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 48.0 }, "checkpoint": "46db4abab3c506113b23b684d6977b1e4e99362fa178f984079e06b0aaf75aa9", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sc3ecd102379d15" ] }, { "age": 90, "six": 45.03333333333333, "five": 47.298, "tasks": { "blimp_filtered": 64.75, "supplement_filtered": 51.56, "ewok_filtered": 51.39, "entity_tracking": 17.83, "comps": 50.96, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 48.0 }, "checkpoint": "7bd9031dcd79bca0ccaa6dda346e8c893189892705cd4daa50e4c6e69ac2c690", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sd1df575c7ef7a5" ] }, { "age": 100, "six": 44.24, "five": 46.558, "tasks": { "blimp_filtered": 64.33, "supplement_filtered": 48.63, "ewok_filtered": 51.08, "entity_tracking": 17.83, "comps": 50.92, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 42.0 }, "checkpoint": "afac36e9892d31530cde8b33f913092d24d7da25181690d9d12a8c5462f41be3", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s8c96ef060becdc" ] } ], "peak": { "five": { "score": 48.12, "ages": [ 10 ], "rank": 1 }, "six": { "score": 45.885, "ages": [ 10 ], "rank": 1 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 225, "id": "legacy-aro20/no_value_basis", "original_name": "旧ARO20M・no_value_basis", "title": "旧ARO20M・no_value_basis", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 47.314166666666665, "five": 49.644, "tasks": { "blimp_filtered": 70.36, "supplement_filtered": 57.26, "ewok_filtered": 51.85, "entity_tracking": 18.01, "comps": 50.74, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 49.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.644, "ages": [ 20 ], "rank": 1 }, "six": { "score": 47.314166666666665, "ages": [ 20 ], "rank": 1 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 226, "id": "legacy-aro20/no_ads+no_value_basis", "original_name": "旧ARO20M・no_ads+no_value_basis", "title": "旧ARO20M・no_ads+no_value_basis", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.983333333333334, "five": 49.564, "tasks": { "blimp_filtered": 69.64, "supplement_filtered": 57.35, "ewok_filtered": 51.97, "entity_tracking": 18.24, "comps": 50.62, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 40.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.564, "ages": [ 20 ], "rank": 2 }, "six": { "score": 46.983333333333334, "ages": [ 20 ], "rank": 2 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 227, "id": "legacy-aro20/no_relation_film+no_value_basis", "original_name": "旧ARO20M・no_relation_film+no_value_basis", "title": "旧ARO20M・no_relation_film+no_value_basis", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.854166666666664, "five": 49.086, "tasks": { "blimp_filtered": 70.59, "supplement_filtered": 57.74, "ewok_filtered": 49.45, "entity_tracking": 17.12, "comps": 50.53, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 51.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.086, "ages": [ 20 ], "rank": 11 }, "six": { "score": 46.854166666666664, "ages": [ 20 ], "rank": 3 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 228, "id": "legacy-aro20/one_deep_pass", "original_name": "旧ARO20M・one_deep_pass", "title": "旧ARO20M・one_deep_pass", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.844166666666666, "five": 49.08, "tasks": { "blimp_filtered": 70.06, "supplement_filtered": 55.63, "ewok_filtered": 51.32, "entity_tracking": 17.46, "comps": 50.93, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 49.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.08, "ages": [ 20 ], "rank": 12 }, "six": { "score": 46.844166666666666, "ages": [ 20 ], "rank": 4 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 229, "id": "legacy-aro20/no_relation_film", "original_name": "旧ARO20M・no_relation_film", "title": "旧ARO20M・no_relation_film", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.7525, "five": 49.464, "tasks": { "blimp_filtered": 70.54, "supplement_filtered": 57.35, "ewok_filtered": 51.06, "entity_tracking": 17.43, "comps": 50.94, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.464, "ages": [ 20 ], "rank": 4 }, "six": { "score": 46.7525, "ages": [ 20 ], "rank": 5 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 230, "id": "legacy-aro20/no_depth", "original_name": "旧ARO20M・no_depth", "title": "旧ARO20M・no_depth", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.705, "five": 49.398, "tasks": { "blimp_filtered": 69.29, "supplement_filtered": 56.52, "ewok_filtered": 52.59, "entity_tracking": 17.73, "comps": 50.86, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 49.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.398, "ages": [ 20 ], "rank": 5 }, "six": { "score": 46.705, "ages": [ 20 ], "rank": 6 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 231, "id": "legacy-aro20/no_value_basis+one_deep_pass", "original_name": "旧ARO20M・no_value_basis+one_deep_pass", "title": "旧ARO20M・no_value_basis+one_deep_pass", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.69583333333333, "five": 49.296, "tasks": { "blimp_filtered": 69.11, "supplement_filtered": 57.21, "ewok_filtered": 51.31, "entity_tracking": 17.96, "comps": 50.89, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.296, "ages": [ 20 ], "rank": 7 }, "six": { "score": 46.69583333333333, "ages": [ 20 ], "rank": 7 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 232, "id": "legacy-aro20/no_future+no_value_basis", "original_name": "旧ARO20M・no_future+no_value_basis", "title": "旧ARO20M・no_future+no_value_basis", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.608333333333334, "five": 49.2, "tasks": { "blimp_filtered": 70.02, "supplement_filtered": 56.51, "ewok_filtered": 51.28, "entity_tracking": 17.67, "comps": 50.52, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 44.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.2, "ages": [ 20 ], "rank": 10 }, "six": { "score": 46.608333333333334, "ages": [ 20 ], "rank": 8 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 233, "id": "legacy-aro20/no_memory+no_value_basis", "original_name": "旧ARO20M・no_memory+no_value_basis", "title": "旧ARO20M・no_memory+no_value_basis", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.56583333333333, "five": 49.046, "tasks": { "blimp_filtered": 69.59, "supplement_filtered": 55.96, "ewok_filtered": 50.52, "entity_tracking": 18.35, "comps": 50.81, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.046, "ages": [ 20 ], "rank": 13 }, "six": { "score": 46.56583333333333, "ages": [ 20 ], "rank": 9 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 234, "id": "legacy-aro20/full", "original_name": "旧ARO20M・full", "title": "旧ARO20M・full", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.54666666666667, "five": 49.506, "tasks": { "blimp_filtered": 70.6, "supplement_filtered": 57.06, "ewok_filtered": 51.23, "entity_tracking": 17.95, "comps": 50.69, "global_piqa_parallel": 16.5, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.506, "ages": [ 20 ], "rank": 3 }, "six": { "score": 46.54666666666667, "ages": [ 20 ], "rank": 10 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 235, "id": "legacy-aro20/no_memory", "original_name": "旧ARO20M・no_memory", "title": "旧ARO20M・no_memory", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.46, "five": 49.304, "tasks": { "blimp_filtered": 69.49, "supplement_filtered": 57.17, "ewok_filtered": 50.37, "entity_tracking": 18.53, "comps": 50.96, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.304, "ages": [ 20 ], "rank": 6 }, "six": { "score": 46.46, "ages": [ 20 ], "rank": 11 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 236, "id": "legacy-aro20/no_ads", "original_name": "旧ARO20M・no_ads", "title": "旧ARO20M・no_ads", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.454166666666666, "five": 49.212, "tasks": { "blimp_filtered": 70.84, "supplement_filtered": 57.51, "ewok_filtered": 49.6, "entity_tracking": 17.56, "comps": 50.55, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 43.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.212, "ages": [ 20 ], "rank": 9 }, "six": { "score": 46.454166666666666, "ages": [ 20 ], "rank": 12 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 237, "id": "legacy-aro20/no_depth+no_value_basis", "original_name": "旧ARO20M・no_depth+no_value_basis", "title": "旧ARO20M・no_depth+no_value_basis", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.454166666666666, "five": 49.012, "tasks": { "blimp_filtered": 69.89, "supplement_filtered": 55.78, "ewok_filtered": 50.24, "entity_tracking": 18.37, "comps": 50.78, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 45.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.012, "ages": [ 20 ], "rank": 14 }, "six": { "score": 46.454166666666666, "ages": [ 20 ], "rank": 12 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 238, "id": "legacy-aro20/no_relation_chain", "original_name": "旧ARO20M・no_relation_chain", "title": "旧ARO20M・no_relation_chain", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.44166666666667, "five": 48.894, "tasks": { "blimp_filtered": 70.06, "supplement_filtered": 55.33, "ewok_filtered": 51.35, "entity_tracking": 17.18, "comps": 50.55, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 48.894, "ages": [ 20 ], "rank": 15 }, "six": { "score": 46.44166666666667, "ages": [ 20 ], "rank": 14 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 239, "id": "legacy-aro20/no_future", "original_name": "旧ARO20M・no_future", "title": "旧ARO20M・no_future", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "legacy-aro20", "cohort_label": "ARO20M", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 46.439166666666665, "five": 49.288, "tasks": { "blimp_filtered": 70.23, "supplement_filtered": 57.91, "ewok_filtered": 49.82, "entity_tracking": 17.88, "comps": 50.6, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 44.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s3b40e0811039cb" ] } ], "peak": { "five": { "score": 49.288, "ages": [ 20 ], "rank": 8 }, "six": { "score": 46.439166666666665, "ages": [ 20 ], "rank": 15 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 240, "id": "run/VISTA_ARO800M_CONTINUE_20260923", "original_name": "VISTA_ARO800M_CONTINUE_20260923", "title": "VISTA_ARO800M_CONTINUE_20260923", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-aro-extended", "cohort_label": "通常データ・ARO延長", "seed": 20260907, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": {}, "tokenizer": {}, "runtime": { "precision": "FP32", "attention": "math_fp32", "tf32": false, "compile": false, "main_threads": 1, "heavy_cpu": "mini", "eval_microbatch": 8, "evaluator_batch_grouping": 64, "eval_forward_microbatch": 8 }, "condition_conflicts": [], "profile": 35, "sources": [ "s548c6a9fbc3d55", "sb18f9c2bb9b77d" ], "points": [ { "age": 610.0, "six": 47.03916666666667, "five": 48.84, "tasks": { "blimp_filtered": 69.4, "supplement_filtered": 54.15, "ewok_filtered": 51.87, "entity_tracking": 16.73, "comps": 52.05, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 45.0 }, "checkpoint": "e0f0424d9d73e85945a8288f0d65ce62931669232f86df5f810d6d4f5df2eac7", "precision": null, "verification": "primary", "sources": [ "s1ade5b8180a562" ] }, { "age": 620.0, "six": 45.700833333333335, "five": 48.214, "tasks": { "blimp_filtered": 68.24, "supplement_filtered": 51.68, "ewok_filtered": 53.5, "entity_tracking": 16.25, "comps": 51.4, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 42.0 }, "checkpoint": "d8c37c34891bd87e1731da14a743877481b9caa2c661fd13c9dcd0a22d0f8b92", "precision": null, "verification": "primary", "sources": [ "s5dfe8678ea9c06" ] }, { "age": 630.0, "six": 47.880833333333335, "five": 49.356, "tasks": { "blimp_filtered": 70.3, "supplement_filtered": 53.2, "ewok_filtered": 52.14, "entity_tracking": 18.86, "comps": 52.28, "global_piqa_parallel": 33.01, "global_piqa_nonparallel": 48.0 }, "checkpoint": "53dc1d8ef198e224d46bfee3fb428238bcc0ae1dd8d3c1fbcf88588000623610", "precision": null, "verification": "primary", "sources": [ "sbc4343f1ddc9d5" ] }, { "age": 640.0, "six": 46.9925, "five": 49.646, "tasks": { "blimp_filtered": 70.56, "supplement_filtered": 53.49, "ewok_filtered": 54.35, "entity_tracking": 17.98, "comps": 51.85, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 49.0 }, "checkpoint": "fe4b9f589783274c63e09e35b64db90193abc881c688011ae6cc9668c15b912b", "precision": null, "verification": "primary", "sources": [ "sdb6ceb5e6a7ee3" ] }, { "age": 650.0, "six": 47.43416666666667, "five": 49.494, "tasks": { "blimp_filtered": 70.11, "supplement_filtered": 55.01, "ewok_filtered": 54.02, "entity_tracking": 16.33, "comps": 52.0, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 50.0 }, "checkpoint": "04ec95692d5905ebb4b5349d5dedca40d8a1f45ec186dfb381cdbe5b273e4d94", "precision": null, "verification": "primary", "sources": [ "sf91241974fe614" ] }, { "age": 660.0, "six": 47.50083333333333, "five": 48.888, "tasks": { "blimp_filtered": 70.54, "supplement_filtered": 52.13, "ewok_filtered": 53.25, "entity_tracking": 16.77, "comps": 51.75, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 52.0 }, "checkpoint": "fe12702a30c621ccf2c7670edb620142da24944b400e9d7a5e2d0ed4b6bd51c4", "precision": null, "verification": "primary", "sources": [ "sdda2d8691f8a02" ] }, { "age": 670.0, "six": 47.638333333333335, "five": 49.05, "tasks": { "blimp_filtered": 70.34, "supplement_filtered": 53.54, "ewok_filtered": 52.89, "entity_tracking": 16.61, "comps": 51.87, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 53.0 }, "checkpoint": "f8b8483917d879ef5af51d9e70e87ed010e18ab6b7a167e1726b1765eda0a876", "precision": null, "verification": "primary", "sources": [ "sf0a5e6a1edf826" ] }, { "age": 680.0, "six": 47.935833333333335, "five": 48.74, "tasks": { "blimp_filtered": 69.57, "supplement_filtered": 52.79, "ewok_filtered": 52.29, "entity_tracking": 17.32, "comps": 51.73, "global_piqa_parallel": 38.83, "global_piqa_nonparallel": 49.0 }, "checkpoint": "8470df09bb08d02e976682183ecb3cfc608929c208f1d850d1d35a9aedbcc6d0", "precision": null, "verification": "primary", "sources": [ "scbcb768f6ae66b" ] }, { "age": 690.0, "six": 47.50666666666667, "five": 49.198, "tasks": { "blimp_filtered": 70.36, "supplement_filtered": 54.22, "ewok_filtered": 52.25, "entity_tracking": 17.51, "comps": 51.65, "global_piqa_parallel": 30.1, "global_piqa_nonparallel": 48.0 }, "checkpoint": "870632f2f4125e3c0609b57493adbb26b379bf50b731eff9271115de8ce883c9", "precision": null, "verification": "primary", "sources": [ "s021c67ab87519a" ] }, { "age": 700.0, "six": 47.48083333333334, "five": 48.97, "tasks": { "blimp_filtered": 69.05, "supplement_filtered": 53.5, "ewok_filtered": 53.29, "entity_tracking": 16.84, "comps": 52.17, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 49.0 }, "checkpoint": "c54202e3c45abff3bee12158de5f4a2e083f3dec7cad874609fa9f6a166cc1ce", "precision": null, "verification": "primary", "sources": [ "s4b66c4356b9600" ] }, { "age": 710.0, "six": 47.053333333333335, "five": 48.854, "tasks": { "blimp_filtered": 69.26, "supplement_filtered": 55.18, "ewok_filtered": 52.41, "entity_tracking": 15.47, "comps": 51.95, "global_piqa_parallel": 30.1, "global_piqa_nonparallel": 46.0 }, "checkpoint": "fa66800c398f28ef611bf0d0fdc888be1e0b01e077e62fcbfd6a87c304923ac7", "precision": null, "verification": "primary", "sources": [ "sf58684fd272381" ] }, { "age": 720.0, "six": 48.24333333333333, "five": 49.3, "tasks": { "blimp_filtered": 70.27, "supplement_filtered": 54.55, "ewok_filtered": 52.84, "entity_tracking": 17.51, "comps": 51.33, "global_piqa_parallel": 35.92, "global_piqa_nonparallel": 50.0 }, "checkpoint": "0f0617d68fae84fcd76880b25c8c0b03973f184dddc13965e762a9bee25a4646", "precision": null, "verification": "primary", "sources": [ "sbd02f1eb56396c" ] }, { "age": 730.0, "six": 47.56, "five": 49.648, "tasks": { "blimp_filtered": 70.41, "supplement_filtered": 55.68, "ewok_filtered": 53.95, "entity_tracking": 16.79, "comps": 51.41, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 49.0 }, "checkpoint": "152569e80cc1ff7cff05ee1a1d4a699b6a1d144a1dfbc04dfbaea27060ebb007", "precision": null, "verification": "primary", "sources": [ "sb48e7b121becb1" ] }, { "age": 740.0, "six": 47.82333333333333, "five": 49.272, "tasks": { "blimp_filtered": 70.56, "supplement_filtered": 52.46, "ewok_filtered": 52.39, "entity_tracking": 19.66, "comps": 51.29, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 53.0 }, "checkpoint": "eed32d601870cdb9a6492d3ff8d1234983583de2a3260e4efecd778fe40987b7", "precision": null, "verification": "primary", "sources": [ "s676fd74308f164" ] }, { "age": 750.0, "six": 46.943333333333335, "five": 48.616, "tasks": { "blimp_filtered": 68.71, "supplement_filtered": 53.88, "ewok_filtered": 52.78, "entity_tracking": 15.84, "comps": 51.87, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 49.0 }, "checkpoint": "5bd4d7f1e0749935355c53b9737e093332fd63304996ba80579381728bf0b01c", "precision": null, "verification": "primary", "sources": [ "s691bd4365c300c" ] }, { "age": 760.0, "six": 48.50416666666667, "five": 49.61, "tasks": { "blimp_filtered": 70.09, "supplement_filtered": 54.68, "ewok_filtered": 54.18, "entity_tracking": 17.59, "comps": 51.51, "global_piqa_parallel": 34.95, "global_piqa_nonparallel": 51.0 }, "checkpoint": "cc7b0714aad504ec36e8fcb0288c5601376f2f6ede45841d1ddefdbc9bd93342", "precision": null, "verification": "primary", "sources": [ "sed188e717d5d9e" ] }, { "age": 770.0, "six": 47.49166666666667, "five": 48.88, "tasks": { "blimp_filtered": 69.33, "supplement_filtered": 53.95, "ewok_filtered": 52.14, "entity_tracking": 17.1, "comps": 51.88, "global_piqa_parallel": 30.1, "global_piqa_nonparallel": 51.0 }, "checkpoint": "8e3653064db7a846867300d4f644988205f660ae9e8a0137b932a5f430048844", "precision": null, "verification": "primary", "sources": [ "s2d1685cc5d3dcc" ] }, { "age": 780.0, "six": 47.13333333333333, "five": 49.03, "tasks": { "blimp_filtered": 69.94, "supplement_filtered": 53.95, "ewok_filtered": 51.97, "entity_tracking": 17.59, "comps": 51.7, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 52.0 }, "checkpoint": "d7efd58003069cb658f5d856fdcbd6fec249e25f86a5deda1e3bb2a4ff55df94", "precision": null, "verification": "primary", "sources": [ "s5a506805b94e08" ] }, { "age": 790.0, "six": 47.340833333333336, "five": 48.796, "tasks": { "blimp_filtered": 69.54, "supplement_filtered": 51.82, "ewok_filtered": 53.1, "entity_tracking": 17.95, "comps": 51.57, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 51.0 }, "checkpoint": "4a397a4d335da774aec8dca816e5f28e17aea4598566ae0733836ea8b8a6e1f5", "precision": null, "verification": "primary", "sources": [ "s95820f7e7f9998" ] }, { "age": 800.0, "six": 47.15, "five": 48.482, "tasks": { "blimp_filtered": 68.23, "supplement_filtered": 51.08, "ewok_filtered": 52.44, "entity_tracking": 18.54, "comps": 52.12, "global_piqa_parallel": 33.98, "global_piqa_nonparallel": 47.0 }, "checkpoint": "dce4bdbfb48fb66aa705dcf2a0156707b5073fc1c5dd17bfe5bbae2afc279cb2", "precision": null, "verification": "primary", "sources": [ "sd8f00806d272ca" ] } ], "peak": { "five": { "score": 49.648, "ages": [ 730.0 ], "rank": 2 }, "six": { "score": 48.50416666666667, "ages": [ 760.0 ], "rank": 1 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 241, "id": "run/VISTA_ARO400M_CONTINUE_20260923", "original_name": "VISTA_ARO400M_CONTINUE_20260923", "title": "VISTA_ARO400M_CONTINUE_20260923", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-aro-extended", "cohort_label": "通常データ・ARO延長", "seed": 20260907, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": {}, "tokenizer": {}, "runtime": { "precision": "FP32", "attention": "math_fp32", "tf32": false, "compile": false, "main_threads": 1, "heavy_cpu": "mini", "eval_microbatch": 8, "evaluator_batch_grouping": 64, "eval_forward_microbatch": 8 }, "condition_conflicts": [], "profile": 35, "sources": [ "s6c07ce2b76f094" ], "points": [ { "age": 210.0, "six": 46.83, "five": 48.278, "tasks": { "blimp_filtered": 68.05, "supplement_filtered": 52.85, "ewok_filtered": 51.66, "entity_tracking": 17.58, "comps": 51.25, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 52.0 }, "checkpoint": "afbb1a70e663b1aef61588cd739594d58a1cd0b4df1e2179ae03350e8ae26073", "precision": null, "verification": "primary", "sources": [ "sa34d10bdfe615d" ] }, { "age": 220.0, "six": 46.76416666666667, "five": 48.396, "tasks": { "blimp_filtered": 68.35, "supplement_filtered": 53.06, "ewok_filtered": 52.01, "entity_tracking": 17.25, "comps": 51.31, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 51.0 }, "checkpoint": "d8f2d76272e5e6df98d03a74d9bbd8f0cd56bbf2333622bc805b5a0f683d8ba7", "precision": null, "verification": "primary", "sources": [ "s2496ef8f75d63f" ] }, { "age": 230.0, "six": 46.306666666666665, "five": 48.65, "tasks": { "blimp_filtered": 67.74, "supplement_filtered": 54.23, "ewok_filtered": 52.22, "entity_tracking": 17.63, "comps": 51.43, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 42.0 }, "checkpoint": "d31f65445936cf148e7272757581fd0cb2241dc0334e9ff5489de1b35f63ad38", "precision": null, "verification": "primary", "sources": [ "s652a542f2d6fd4" ] }, { "age": 240.0, "six": 46.6325, "five": 48.246, "tasks": { "blimp_filtered": 67.02, "supplement_filtered": 53.5, "ewok_filtered": 52.0, "entity_tracking": 17.7, "comps": 51.01, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 48.0 }, "checkpoint": "6eb1148ae7d600ef34871af1f4a7be9259a2fa3717ea90cd95db42aeabe7b5bf", "precision": null, "verification": "primary", "sources": [ "scc5407f5b78cbb" ] }, { "age": 250.0, "six": 46.75416666666667, "five": 48.566, "tasks": { "blimp_filtered": 68.73, "supplement_filtered": 54.55, "ewok_filtered": 51.96, "entity_tracking": 16.14, "comps": 51.45, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 55.0 }, "checkpoint": "d71c6054c4e2b8b1b8e7f320f74d0e16cd077d8c9ff66f103eae16907fff5baf", "precision": null, "verification": "primary", "sources": [ "sdb3a8c35e94e99" ] }, { "age": 260.0, "six": 47.09166666666667, "five": 48.6, "tasks": { "blimp_filtered": 68.96, "supplement_filtered": 53.64, "ewok_filtered": 52.55, "entity_tracking": 16.53, "comps": 51.32, "global_piqa_parallel": 30.1, "global_piqa_nonparallel": 49.0 }, "checkpoint": "b689b3bf748fae7433dc29b3ce2b113bee3684d47b8dda3d43b32f6ac9f3c73b", "precision": null, "verification": "primary", "sources": [ "s8b39da87cf622e" ] }, { "age": 270.0, "six": 47.4975, "five": 48.902, "tasks": { "blimp_filtered": 67.97, "supplement_filtered": 54.81, "ewok_filtered": 52.33, "entity_tracking": 17.83, "comps": 51.57, "global_piqa_parallel": 34.95, "global_piqa_nonparallel": 46.0 }, "checkpoint": "72a95b7aaac233d23d5316bcceb9e53d4888a4a4f4ff8b80326cc8f5374449be", "precision": null, "verification": "primary", "sources": [ "s006ed0fc395fd4" ] }, { "age": 280.0, "six": 47.653333333333336, "five": 48.874, "tasks": { "blimp_filtered": 68.6, "supplement_filtered": 54.05, "ewok_filtered": 52.22, "entity_tracking": 17.89, "comps": 51.61, "global_piqa_parallel": 30.1, "global_piqa_nonparallel": 53.0 }, "checkpoint": "29f2b7e3f48e16e759367763c512f3e5bfb3755516139281a029c15ef57dd7a8", "precision": null, "verification": "primary", "sources": [ "sc63595bc3898ea" ] }, { "age": 290.0, "six": 47.405, "five": 49.87, "tasks": { "blimp_filtered": 69.81, "supplement_filtered": 55.33, "ewok_filtered": 54.07, "entity_tracking": 18.6, "comps": 51.54, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 42.0 }, "checkpoint": "02b49defc49d7dca37d2e826486e0af8fc58c987b45db4c774b7051165a10608", "precision": null, "verification": "primary", "sources": [ "sa4eca564e6a24a" ] }, { "age": 300.0, "six": 46.7075, "five": 48.736, "tasks": { "blimp_filtered": 67.97, "supplement_filtered": 53.52, "ewok_filtered": 51.67, "entity_tracking": 19.2, "comps": 51.32, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 44.0 }, "checkpoint": "848a26ad7380ad14646eb03562faa5848f37fc8c9a7f886c0419ebca64acc27d", "precision": null, "verification": "primary", "sources": [ "s1915be4af955e3" ] }, { "age": 310.0, "six": 46.06583333333333, "five": 47.858, "tasks": { "blimp_filtered": 67.57, "supplement_filtered": 50.98, "ewok_filtered": 52.11, "entity_tracking": 17.39, "comps": 51.24, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 48.0 }, "checkpoint": "0b9b1d4e7ef56cfb1e2a945f291409c330f5adf0da6dcc0d4957ccc81ec542b2", "precision": null, "verification": "primary", "sources": [ "s31442e6480b0ac" ] }, { "age": 320.0, "six": 46.925, "five": 48.386, "tasks": { "blimp_filtered": 69.49, "supplement_filtered": 51.53, "ewok_filtered": 52.42, "entity_tracking": 16.63, "comps": 51.86, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 54.0 }, "checkpoint": "ab433f9e0fefd85ce62d89e333e94e91e16349bb44c8694aa10968df2e37b52d", "precision": null, "verification": "primary", "sources": [ "s978e894541e19b" ] }, { "age": 330.0, "six": 47.5375, "five": 49.332, "tasks": { "blimp_filtered": 69.22, "supplement_filtered": 53.32, "ewok_filtered": 52.65, "entity_tracking": 19.99, "comps": 51.48, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 48.0 }, "checkpoint": "af29f0ef6ad199f51492e924fcbe913781cfe5046ee74f57148bbb20e879c10d", "precision": null, "verification": "primary", "sources": [ "s4b17eb4dc9db02" ] }, { "age": 340.0, "six": 46.98583333333333, "five": 48.462, "tasks": { "blimp_filtered": 69.69, "supplement_filtered": 51.04, "ewok_filtered": 51.91, "entity_tracking": 18.67, "comps": 51.0, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 53.0 }, "checkpoint": "8716adf49c22c8b137e1f3effc177c740a5ae4b0859e02321bbc497c233426d1", "precision": null, "verification": "primary", "sources": [ "s066e58fec7bf96" ] }, { "age": 350.0, "six": 47.47, "five": 49.734, "tasks": { "blimp_filtered": 71.04, "supplement_filtered": 55.45, "ewok_filtered": 52.43, "entity_tracking": 18.0, "comps": 51.75, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 49.0 }, "checkpoint": "0e707518e406d75c5146b63f99826da19005ed05f3d1c78a0a77076c8d0f8c5b", "precision": null, "verification": "primary", "sources": [ "s2d12a4631d8510" ] }, { "age": 360.0, "six": 47.8875, "five": 49.358, "tasks": { "blimp_filtered": 70.51, "supplement_filtered": 53.3, "ewok_filtered": 52.91, "entity_tracking": 17.83, "comps": 52.24, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 50.0 }, "checkpoint": "6a2c75a79885a188066dd1498ca0d76b6331c4092cd91c8d9f4b212d087d2af0", "precision": null, "verification": "primary", "sources": [ "s75e21f4051d1e6" ] }, { "age": 370.0, "six": 47.0375, "five": 48.824, "tasks": { "blimp_filtered": 69.93, "supplement_filtered": 53.42, "ewok_filtered": 52.17, "entity_tracking": 16.78, "comps": 51.82, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 50.0 }, "checkpoint": "4a63a4b11ed32845ec3e522aec0fe29f614612a77db3a1a480eb6bc20deeee23", "precision": null, "verification": "primary", "sources": [ "s8efe17bb9af0a1" ] }, { "age": 380.0, "six": 47.1725, "five": 48.5, "tasks": { "blimp_filtered": 70.32, "supplement_filtered": 50.2, "ewok_filtered": 52.14, "entity_tracking": 18.37, "comps": 51.47, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 50.0 }, "checkpoint": "5979b86ed8adff459a8a74c9940730c62458ae25935d65505db8e9bf2d9b7c7a", "precision": null, "verification": "primary", "sources": [ "sf6f1451a1bf077" ] }, { "age": 390.0, "six": 46.98166666666667, "five": 48.76, "tasks": { "blimp_filtered": 68.82, "supplement_filtered": 52.08, "ewok_filtered": 52.04, "entity_tracking": 19.55, "comps": 51.31, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 49.0 }, "checkpoint": "cb26112a62dba5ca259d5320667e0905ad0e57f47fdbe62336750e40ebfb69ec", "precision": null, "verification": "primary", "sources": [ "sb89a13d826bbe7" ] }, { "age": 400.0, "six": 47.0275, "five": 49.006, "tasks": { "blimp_filtered": 70.18, "supplement_filtered": 55.43, "ewok_filtered": 50.65, "entity_tracking": 17.16, "comps": 51.61, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 50.0 }, "checkpoint": "ad0856360b0e73874e15cdbf7c1f4b1e967569fa2e4605db2f61e4acb0f92bfd", "precision": null, "verification": "primary", "sources": [ "s77e83cc0588c2e" ] } ], "peak": { "five": { "score": 49.87, "ages": [ 290.0 ], "rank": 1 }, "six": { "score": 47.8875, "ages": [ 360.0 ], "rank": 2 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 242, "id": "run/VISTA_ARO600M_CONTINUE_20260923", "original_name": "VISTA_ARO600M_CONTINUE_20260923", "title": "VISTA_ARO600M_CONTINUE_20260923", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-aro-extended", "cohort_label": "通常データ・ARO延長", "seed": 20260907, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": {}, "tokenizer": {}, "runtime": { "precision": "FP32", "attention": "math_fp32", "tf32": false, "compile": false, "main_threads": 1, "heavy_cpu": "mini", "eval_microbatch": 8, "evaluator_batch_grouping": 64, "eval_forward_microbatch": 8 }, "condition_conflicts": [], "profile": 35, "sources": [ "secf9fbc20fe448", "s387658e8b71571" ], "points": [ { "age": 510.0, "six": 47.119166666666665, "five": 48.622, "tasks": { "blimp_filtered": 69.02, "supplement_filtered": 52.31, "ewok_filtered": 52.36, "entity_tracking": 17.73, "comps": 51.69, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 53.0 }, "checkpoint": "4cf3ef1e5ba42cd543e03fb51c9b84599d33698b4f3fe9410789931dd2b718c4", "precision": null, "verification": "primary", "sources": [ "s21690d259acb21" ] }, { "age": 520.0, "six": 46.30166666666667, "five": 48.144, "tasks": { "blimp_filtered": 70.27, "supplement_filtered": 49.96, "ewok_filtered": 52.61, "entity_tracking": 16.32, "comps": 51.56, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 47.0 }, "checkpoint": "19170ebb5cde2e1303c35f7777228a597d0fb3003e02b43686a45e10690548b4", "precision": null, "verification": "primary", "sources": [ "s9648b892cfb613" ] }, { "age": 530.0, "six": 46.62583333333333, "five": 48.138, "tasks": { "blimp_filtered": 69.1, "supplement_filtered": 50.68, "ewok_filtered": 51.92, "entity_tracking": 17.51, "comps": 51.48, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 49.0 }, "checkpoint": "d2c3acdaefb5662579b03e618b46db51d170830ba007e604534d91566abc798b", "precision": null, "verification": "primary", "sources": [ "sb1490b89f7df56" ] }, { "age": 540.0, "six": 47.431666666666665, "five": 49.1, "tasks": { "blimp_filtered": 68.75, "supplement_filtered": 52.74, "ewok_filtered": 54.01, "entity_tracking": 18.1, "comps": 51.9, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 51.0 }, "checkpoint": "0cb9fbdc764dfbfb1738291325e8113847a22d095e8c56690befe860fdb6ee12", "precision": null, "verification": "primary", "sources": [ "s1152ca6363f31c" ] }, { "age": 550.0, "six": 47.675, "five": 49.394, "tasks": { "blimp_filtered": 68.42, "supplement_filtered": 54.55, "ewok_filtered": 52.0, "entity_tracking": 20.36, "comps": 51.64, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 50.0 }, "checkpoint": "1247f0a597b0f74232672038b09ef21e59c5189baa9de99fa8a9430033ae4754", "precision": null, "verification": "primary", "sources": [ "s7393720d8fc517" ] }, { "age": 560.0, "six": 46.62583333333333, "five": 48.83, "tasks": { "blimp_filtered": 69.42, "supplement_filtered": 52.41, "ewok_filtered": 51.03, "entity_tracking": 19.7, "comps": 51.59, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 45.0 }, "checkpoint": "7a1d7a53b3e1a7248e4bb23ada801ca4fbbcb0fe3e9c17ca12cac458305a7f85", "precision": null, "verification": "primary", "sources": [ "se3e6214a2d51fb" ] }, { "age": 570.0, "six": 47.229166666666664, "five": 49.454, "tasks": { "blimp_filtered": 70.08, "supplement_filtered": 54.3, "ewok_filtered": 52.84, "entity_tracking": 18.65, "comps": 51.4, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 46.0 }, "checkpoint": "06bfea7ffab9e314996ff4435c083b8f44815d5bc46ea7fe04541d37551dfcdf", "precision": null, "verification": "primary", "sources": [ "s6a7c6ae75796ae" ] }, { "age": 580.0, "six": 47.545, "five": 49.536, "tasks": { "blimp_filtered": 69.59, "supplement_filtered": 53.66, "ewok_filtered": 52.96, "entity_tracking": 19.46, "comps": 52.01, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 48.0 }, "checkpoint": "64d09e3011f81cbcd9377baef5ffe333ad5c4948a40476772f90c67f168d709f", "precision": null, "verification": "primary", "sources": [ "s4677e7e1764e94" ] }, { "age": 590.0, "six": 47.25833333333333, "five": 49.3, "tasks": { "blimp_filtered": 69.53, "supplement_filtered": 52.45, "ewok_filtered": 55.25, "entity_tracking": 17.61, "comps": 51.66, "global_piqa_parallel": 30.1, "global_piqa_nonparallel": 44.0 }, "checkpoint": "e2b896023c958cae08a25918b497e639306d9f01df5b1091a8bc6c8715a96a9f", "precision": null, "verification": "primary", "sources": [ "s3f49cdbd2dd797" ] }, { "age": 600.0, "six": 47.615, "five": 48.834, "tasks": { "blimp_filtered": 68.89, "supplement_filtered": 51.54, "ewok_filtered": 53.45, "entity_tracking": 18.61, "comps": 51.68, "global_piqa_parallel": 32.04, "global_piqa_nonparallel": 51.0 }, "checkpoint": "2542723d7b3129783a18e5c77a335f4eeedc67f1e6cfa976ccb935b4ffcfe4e1", "precision": null, "verification": "primary", "sources": [ "s1f58123caf30d6" ] } ], "peak": { "five": { "score": 49.536, "ages": [ 580.0 ], "rank": 4 }, "six": { "score": 47.675, "ages": [ 550.0 ], "rank": 3 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 243, "id": "run/VISTA_ARO200M_CONTINUE_20260922", "original_name": "VISTA_ARO200M_CONTINUE_20260922", "title": "VISTA_ARO200M_CONTINUE_20260922", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-aro-extended", "cohort_label": "通常データ・ARO延長", "seed": 20260907, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": {}, "tokenizer": {}, "runtime": { "precision": "FP32", "attention": "math_fp32", "tf32": false, "compile": false, "main_threads": 1, "heavy_cpu": "mini", "eval_microbatch": 8, "evaluator_batch_grouping": 64, "eval_forward_microbatch": 8 }, "condition_conflicts": [], "profile": 35, "sources": [ "s3f401bd189bad6" ], "points": [ { "age": 110.0, "six": 46.7425, "five": 48.478, "tasks": { "blimp_filtered": 68.59, "supplement_filtered": 51.35, "ewok_filtered": 51.58, "entity_tracking": 19.55, "comps": 51.32, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 47.0 }, "checkpoint": "392f6d471f2ec8508053b61321bef536fbc4d0d2e71381fb77c9290370fe2951", "precision": null, "verification": "primary", "sources": [ "s90edbd86d8498d" ] }, { "age": 120.0, "six": 46.76916666666666, "five": 48.802, "tasks": { "blimp_filtered": 68.1, "supplement_filtered": 53.81, "ewok_filtered": 51.6, "entity_tracking": 18.99, "comps": 51.51, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 47.0 }, "checkpoint": "e66669dd23f31a1e9668f77eeb89af970b129255f729a5ac9adf4583a4ad68c9", "precision": null, "verification": "primary", "sources": [ "s995913c772857d" ] }, { "age": 130.0, "six": 47.3675, "five": 49.028, "tasks": { "blimp_filtered": 67.32, "supplement_filtered": 53.9, "ewok_filtered": 51.33, "entity_tracking": 20.93, "comps": 51.66, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 49.0 }, "checkpoint": "09b3a2d447f7a7373ab6233e974ffc75f14044341cda32777f90417daa1ee8b3", "precision": null, "verification": "primary", "sources": [ "sc89cf0b6666601" ] }, { "age": 140.0, "six": 46.2525, "five": 48.19, "tasks": { "blimp_filtered": 67.97, "supplement_filtered": 51.53, "ewok_filtered": 51.81, "entity_tracking": 17.86, "comps": 51.78, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 44.0 }, "checkpoint": "48aeebd3d9eb5521414b23f3202c229c89cf789bd6c257bb3caf7c8b06ac965e", "precision": null, "verification": "primary", "sources": [ "sadd1ce5bfdea21" ] }, { "age": 150.0, "six": 47.498333333333335, "five": 49.58, "tasks": { "blimp_filtered": 68.43, "supplement_filtered": 56.05, "ewok_filtered": 52.65, "entity_tracking": 19.31, "comps": 51.46, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 47.0 }, "checkpoint": "6316bc2aea2a649a76a107b0a596bba5f744563f509f71d297c7490b235d519d", "precision": null, "verification": "primary", "sources": [ "s07d177c110029b" ] }, { "age": 160.0, "six": 46.6975, "five": 49.016, "tasks": { "blimp_filtered": 67.32, "supplement_filtered": 56.55, "ewok_filtered": 51.89, "entity_tracking": 18.27, "comps": 51.05, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 44.0 }, "checkpoint": "36030a1f8612d1ce84221e74c900a851957c33187d2b46f0fcf91b682b2985a8", "precision": null, "verification": "primary", "sources": [ "s117ce8f145368e" ] }, { "age": 170.0, "six": 46.38916666666667, "five": 48.74, "tasks": { "blimp_filtered": 67.51, "supplement_filtered": 54.77, "ewok_filtered": 52.57, "entity_tracking": 17.64, "comps": 51.21, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 45.0 }, "checkpoint": "1f5ae26b086706537c91a5492b0fb2d5fd1f768afed71c089131920186e4e9f8", "precision": null, "verification": "primary", "sources": [ "sa068bfc07d99bb" ] }, { "age": 180.0, "six": 46.9775, "five": 48.76, "tasks": { "blimp_filtered": 68.65, "supplement_filtered": 52.5, "ewok_filtered": 52.45, "entity_tracking": 18.91, "comps": 51.29, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 47.0 }, "checkpoint": "bbc438243ae65a51f56d357ec253dc0640d49e2c5e129eec49feca5b65403761", "precision": null, "verification": "primary", "sources": [ "s6ca3d6cd6408b9" ] }, { "age": 190.0, "six": 46.4175, "five": 48.388, "tasks": { "blimp_filtered": 67.96, "supplement_filtered": 52.18, "ewok_filtered": 52.12, "entity_tracking": 18.43, "comps": 51.25, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 44.0 }, "checkpoint": "24499f7da502ef6ac6595eaa907a10957105c0cbb13053e8568b2fd9217429a2", "precision": null, "verification": "primary", "sources": [ "s25ee6484ccaee0" ] }, { "age": 200.0, "six": 45.75, "five": 48.09, "tasks": { "blimp_filtered": 69.22, "supplement_filtered": 50.71, "ewok_filtered": 51.24, "entity_tracking": 17.42, "comps": 51.86, "global_piqa_parallel": 30.1, "global_piqa_nonparallel": 38.0 }, "checkpoint": "d652abaa8d898561bf12fa8ef16a90e61038f1489bb9d942c9bbdea72f7adc70", "precision": null, "verification": "primary", "sources": [ "sa604e14ba30a98" ] } ], "peak": { "five": { "score": 49.58, "ages": [ 150.0 ], "rank": 3 }, "six": { "score": 47.498333333333335, "ages": [ 150.0 ], "rank": 4 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 244, "id": "run/VISTA_ARO500M_CONTINUE_20260923", "original_name": "VISTA_ARO500M_CONTINUE_20260923", "title": "VISTA_ARO500M_CONTINUE_20260923", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-aro-extended", "cohort_label": "通常データ・ARO延長", "seed": 20260907, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": {}, "tokenizer": {}, "runtime": { "precision": "FP32", "attention": "math_fp32", "tf32": false, "compile": false, "main_threads": 1, "heavy_cpu": "mini", "eval_microbatch": 8, "evaluator_batch_grouping": 64, "eval_forward_microbatch": 8 }, "condition_conflicts": [], "profile": 35, "sources": [ "s5cd50c55c9fc25" ], "points": [ { "age": 410.0, "six": 47.166666666666664, "five": 48.884, "tasks": { "blimp_filtered": 68.26, "supplement_filtered": 52.68, "ewok_filtered": 52.98, "entity_tracking": 18.82, "comps": 51.68, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 49.0 }, "checkpoint": "e576b73d19a0c95b43606517ab2303e0904d1cde6678b2baacea898586b34389", "precision": null, "verification": "primary", "sources": [ "s3f45b730c14a8a" ] }, { "age": 420.0, "six": 47.285833333333336, "five": 49.33, "tasks": { "blimp_filtered": 69.97, "supplement_filtered": 53.35, "ewok_filtered": 54.42, "entity_tracking": 17.55, "comps": 51.36, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 45.0 }, "checkpoint": "4ddbb0a64102dca689d6e5dc4445e462ddda09e9303b117f5c57ba04806302c5", "precision": null, "verification": "primary", "sources": [ "sab1726fb6bce7b" ] }, { "age": 430.0, "six": 46.788333333333334, "five": 48.728, "tasks": { "blimp_filtered": 69.18, "supplement_filtered": 51.65, "ewok_filtered": 52.54, "entity_tracking": 18.72, "comps": 51.55, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 47.0 }, "checkpoint": "5223000e151bd92c28b5ae9db29a844a38e09ba42e08187bebb03ef982048c3f", "precision": null, "verification": "primary", "sources": [ "s36871fae1d2974" ] }, { "age": 440.0, "six": 47.09916666666667, "five": 49.498, "tasks": { "blimp_filtered": 70.9, "supplement_filtered": 52.65, "ewok_filtered": 52.69, "entity_tracking": 19.45, "comps": 51.8, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 44.0 }, "checkpoint": "53b54845c0ecebe9624d866205fde259543ecd2a79b33a935fb6ee8947646552", "precision": null, "verification": "primary", "sources": [ "sfc25339d609421" ] }, { "age": 450.0, "six": 46.835, "five": 48.884, "tasks": { "blimp_filtered": 69.44, "supplement_filtered": 52.29, "ewok_filtered": 51.88, "entity_tracking": 19.37, "comps": 51.44, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 46.0 }, "checkpoint": "c033de5c3c5485717f9358bfba89d3ad28570784ad80a401d4a169a5949e22ab", "precision": null, "verification": "primary", "sources": [ "s4e49f26adf80a1" ] }, { "age": 460.0, "six": 46.50333333333333, "five": 48.68, "tasks": { "blimp_filtered": 69.36, "supplement_filtered": 53.5, "ewok_filtered": 52.26, "entity_tracking": 16.94, "comps": 51.34, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 46.0 }, "checkpoint": "045eee3196de613dfae2617244a566f33dccda45ae5ddb53b94bb0c9400ec084", "precision": null, "verification": "primary", "sources": [ "s794d5e1aeba882" ] }, { "age": 470.0, "six": 46.81333333333333, "five": 48.76, "tasks": { "blimp_filtered": 66.84, "supplement_filtered": 54.36, "ewok_filtered": 52.2, "entity_tracking": 18.88, "comps": 51.52, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 46.0 }, "checkpoint": "39e4627912261d7d7bb93356cf5dd9a879af514939c316d5974b643b0ea012f3", "precision": null, "verification": "primary", "sources": [ "scd25f95bebcf59" ] }, { "age": 480.0, "six": 46.375, "five": 48.432, "tasks": { "blimp_filtered": 69.6, "supplement_filtered": 52.56, "ewok_filtered": 52.15, "entity_tracking": 16.46, "comps": 51.39, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 45.0 }, "checkpoint": "022289cdc3ce564a8c543cd13a43c5171ce9ec1310d52f6448ca71ffce6295fe", "precision": null, "verification": "primary", "sources": [ "s2994191a10693a" ] }, { "age": 490.0, "six": 46.464166666666664, "five": 48.924, "tasks": { "blimp_filtered": 68.72, "supplement_filtered": 53.74, "ewok_filtered": 53.95, "entity_tracking": 16.7, "comps": 51.51, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 46.0 }, "checkpoint": "d66b8e3da8a73be5ecdb81a1ae30acb343e2ab11bd77715103fa2d6e2323073b", "precision": null, "verification": "primary", "sources": [ "s032b5315ed44fb" ] }, { "age": 500.0, "six": 47.30166666666667, "five": 48.946, "tasks": { "blimp_filtered": 69.85, "supplement_filtered": 53.46, "ewok_filtered": 53.47, "entity_tracking": 16.57, "comps": 51.38, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 50.0 }, "checkpoint": "8729194528fa2ff42dbb16352807bf026046d2ddf2ada4fed427af15049020d4", "precision": null, "verification": "primary", "sources": [ "sad1826666b806e" ] } ], "peak": { "five": { "score": 49.498, "ages": [ 440.0 ], "rank": 5 }, "six": { "score": 47.30166666666667, "ages": [ 500.0 ], "rank": 5 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 245, "id": "run/VISTA_ARO100M_SCRATCH_20260922", "original_name": "VISTA_ARO100M_SCRATCH_20260922", "title": "VISTA_ARO100M_SCRATCH_20260922", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-aro-extended", "cohort_label": "通常データ・ARO延長", "seed": 20260907, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": {}, "tokenizer": {}, "runtime": { "precision": "FP32", "attention": "math_fp32", "tf32": false, "compile": false, "main_threads": 1, "heavy_cpu": "mini", "eval_microbatch": 8, "evaluator_batch_grouping": 64, "eval_forward_microbatch": 8 }, "condition_conflicts": [], "profile": 35, "sources": [ "sb57960a5684617" ], "points": [ { "age": 10.0, "six": 43.026666666666664, "five": 45.19, "tasks": { "blimp_filtered": 60.42, "supplement_filtered": 47.96, "ewok_filtered": 51.07, "entity_tracking": 16.39, "comps": 50.11, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 45.0 }, "checkpoint": "c848fffc677ec23dbc0e4e8f2620d37163e960b248fa3e18ca4de16a0c83673e", "precision": null, "verification": "primary", "sources": [ "s6b027cabc36c72" ] }, { "age": 20.0, "six": 44.69083333333333, "five": 45.822, "tasks": { "blimp_filtered": 62.31, "supplement_filtered": 48.72, "ewok_filtered": 50.38, "entity_tracking": 17.98, "comps": 49.72, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 47.0 }, "checkpoint": "e517e288cde0dd874c4e362f6b69e4599440cdabcffa401bee1e218c63858469", "precision": null, "verification": "primary", "sources": [ "sf4956fb3200119" ] }, { "age": 30.0, "six": 45.325833333333335, "five": 46.77, "tasks": { "blimp_filtered": 62.96, "supplement_filtered": 52.7, "ewok_filtered": 50.02, "entity_tracking": 17.74, "comps": 50.43, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 50.0 }, "checkpoint": "3b6fb0a45b82f8d3f71021d6e0bac3c282a696a2d055ecb0269801b7e204ba9a", "precision": null, "verification": "primary", "sources": [ "s64aa42a8ad3aa1" ] }, { "age": 40.0, "six": 45.64, "five": 47.038, "tasks": { "blimp_filtered": 64.44, "supplement_filtered": 52.78, "ewok_filtered": 49.61, "entity_tracking": 18.41, "comps": 49.95, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 54.0 }, "checkpoint": "1646bebef8058649b0018dbce0480ecc089ccf438f509ed925a7e6f3bc624143", "precision": null, "verification": "primary", "sources": [ "s54ae89dfff9604" ] }, { "age": 50.0, "six": 45.5025, "five": 47.67, "tasks": { "blimp_filtered": 64.93, "supplement_filtered": 52.04, "ewok_filtered": 51.83, "entity_tracking": 19.34, "comps": 50.21, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "524a3a3a104d38764e0cff4630ae4dc80646358f5e5cbff382ba9adbe89709a1", "precision": null, "verification": "primary", "sources": [ "s47576f67089bb3" ] }, { "age": 60.0, "six": 45.795, "five": 48.03, "tasks": { "blimp_filtered": 66.63, "supplement_filtered": 51.83, "ewok_filtered": 51.1, "entity_tracking": 20.45, "comps": 50.14, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 44.0 }, "checkpoint": "eda02de3f31e1b8405e6838d2dcf3ddc3e85193d4f2b12935a6173b3452a7340", "precision": null, "verification": "primary", "sources": [ "s4201f95c26b92d" ] }, { "age": 70.0, "six": 46.45333333333333, "five": 48.028, "tasks": { "blimp_filtered": 65.85, "supplement_filtered": 54.17, "ewok_filtered": 51.36, "entity_tracking": 18.27, "comps": 50.49, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 49.0 }, "checkpoint": "ff3b0db010c78d8f1b0909f20373eae3c5a1ec85e619a2dafd196019a073b03a", "precision": null, "verification": "primary", "sources": [ "s7be9646d3810d6" ] }, { "age": 80.0, "six": 44.96, "five": 47.216, "tasks": { "blimp_filtered": 66.68, "supplement_filtered": 48.7, "ewok_filtered": 50.28, "entity_tracking": 19.07, "comps": 51.35, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 46.0 }, "checkpoint": "e6afecfd9ba5dddc0f5e99427530acb488aca31e8bb960b2d4dca12639c0236e", "precision": null, "verification": "primary", "sources": [ "s007254ea824950" ] }, { "age": 90.0, "six": 46.263333333333335, "five": 47.8, "tasks": { "blimp_filtered": 66.87, "supplement_filtered": 50.69, "ewok_filtered": 50.14, "entity_tracking": 20.33, "comps": 50.97, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 49.0 }, "checkpoint": "dfe193c8df218a18861fe8281d7ddefaf3bdcfe0f4df2f1b2f74681276b1a73c", "precision": null, "verification": "primary", "sources": [ "s3e202338d52ecb" ] } ], "peak": { "five": { "score": 48.03, "ages": [ 60.0 ], "rank": 6 }, "six": { "score": 46.45333333333333, "ages": [ 70.0 ], "rank": 6 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 246, "id": "normal-joint/joint_04", "original_name": "joint_04", "title": "joint_04", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-joint", "cohort_label": "通常データ・joint探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 45.06916666666667, "five": 46.656, "tasks": { "blimp_filtered": 62.61, "supplement_filtered": 51.63, "ewok_filtered": 50.68, "entity_tracking": 18.12, "comps": 50.24, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 50.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s66f4e4e5536aa7" ] }, { "age": 60, "six": 45.979166666666664, "five": 47.368, "tasks": { "blimp_filtered": 66.25, "supplement_filtered": 51.26, "ewok_filtered": 50.69, "entity_tracking": 18.63, "comps": 50.01, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 70, "six": 46.23, "five": 47.466, "tasks": { "blimp_filtered": 66.51, "supplement_filtered": 51.86, "ewok_filtered": 49.82, "entity_tracking": 18.76, "comps": 50.38, "global_piqa_parallel": 30.1, "global_piqa_nonparallel": 50.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 80, "six": 46.1275, "five": 47.532, "tasks": { "blimp_filtered": 66.28, "supplement_filtered": 49.1, "ewok_filtered": 51.68, "entity_tracking": 19.54, "comps": 51.06, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 52.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 90, "six": 45.49666666666667, "five": 47.378, "tasks": { "blimp_filtered": 66.72, "supplement_filtered": 47.93, "ewok_filtered": 52.0, "entity_tracking": 19.34, "comps": 50.9, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 45.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 100, "six": 46.865833333333335, "five": 48.432, "tasks": { "blimp_filtered": 68.5, "supplement_filtered": 51.43, "ewok_filtered": 52.64, "entity_tracking": 18.57, "comps": 51.02, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] } ], "peak": { "five": { "score": 48.432, "ages": [ 100 ], "rank": 3 }, "six": { "score": 46.865833333333335, "ages": [ 100 ], "rank": 1 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 247, "id": "normal-joint/joint_09", "original_name": "joint_09", "title": "joint_09", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-joint", "cohort_label": "通常データ・joint探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 45.36833333333333, "five": 46.232, "tasks": { "blimp_filtered": 62.89, "supplement_filtered": 48.81, "ewok_filtered": 50.41, "entity_tracking": 19.07, "comps": 49.98, "global_piqa_parallel": 30.1, "global_piqa_nonparallel": 52.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s66f4e4e5536aa7" ] }, { "age": 60, "six": 46.0975, "five": 47.81, "tasks": { "blimp_filtered": 65.36, "supplement_filtered": 51.52, "ewok_filtered": 52.22, "entity_tracking": 19.32, "comps": 50.63, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 44.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 70, "six": 46.46666666666667, "five": 47.85, "tasks": { "blimp_filtered": 65.14, "supplement_filtered": 53.95, "ewok_filtered": 50.74, "entity_tracking": 18.92, "comps": 50.5, "global_piqa_parallel": 30.1, "global_piqa_nonparallel": 49.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 80, "six": 46.33083333333333, "five": 48.47, "tasks": { "blimp_filtered": 66.73, "supplement_filtered": 53.69, "ewok_filtered": 50.82, "entity_tracking": 19.77, "comps": 51.34, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 90, "six": 46.55416666666667, "five": 48.444, "tasks": { "blimp_filtered": 66.87, "supplement_filtered": 51.17, "ewok_filtered": 51.87, "entity_tracking": 21.01, "comps": 51.3, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 100, "six": 46.788333333333334, "five": 48.73, "tasks": { "blimp_filtered": 67.92, "supplement_filtered": 51.36, "ewok_filtered": 52.71, "entity_tracking": 20.26, "comps": 51.4, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] } ], "peak": { "five": { "score": 48.73, "ages": [ 100 ], "rank": 1 }, "six": { "score": 46.788333333333334, "ages": [ 100 ], "rank": 2 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 248, "id": "normal-joint/joint_08", "original_name": "joint_08", "title": "joint_08", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-joint", "cohort_label": "通常データ・joint探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 45.83833333333333, "five": 46.408, "tasks": { "blimp_filtered": 62.95, "supplement_filtered": 50.81, "ewok_filtered": 52.0, "entity_tracking": 16.71, "comps": 49.57, "global_piqa_parallel": 33.98, "global_piqa_nonparallel": 52.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s66f4e4e5536aa7" ] }, { "age": 60, "six": 45.43833333333333, "five": 47.202, "tasks": { "blimp_filtered": 64.94, "supplement_filtered": 50.97, "ewok_filtered": 49.67, "entity_tracking": 19.68, "comps": 50.75, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 70, "six": 46.215833333333336, "five": 47.932, "tasks": { "blimp_filtered": 65.1, "supplement_filtered": 53.65, "ewok_filtered": 50.43, "entity_tracking": 19.48, "comps": 51.0, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 51.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 80, "six": 46.0925, "five": 47.778, "tasks": { "blimp_filtered": 66.68, "supplement_filtered": 51.0, "ewok_filtered": 51.06, "entity_tracking": 18.78, "comps": 51.37, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 53.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 90, "six": 46.05833333333333, "five": 48.246, "tasks": { "blimp_filtered": 65.41, "supplement_filtered": 51.74, "ewok_filtered": 51.72, "entity_tracking": 21.28, "comps": 51.08, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 45.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 100, "six": 46.68333333333333, "five": 48.504, "tasks": { "blimp_filtered": 68.51, "supplement_filtered": 49.76, "ewok_filtered": 53.16, "entity_tracking": 19.53, "comps": 51.56, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] } ], "peak": { "five": { "score": 48.504, "ages": [ 100 ], "rank": 2 }, "six": { "score": 46.68333333333333, "ages": [ 100 ], "rank": 3 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 249, "id": "normal-joint/joint_10", "original_name": "joint_10", "title": "joint_10", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-joint", "cohort_label": "通常データ・joint探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 44.818333333333335, "five": 46.264, "tasks": { "blimp_filtered": 64.43, "supplement_filtered": 49.73, "ewok_filtered": 49.95, "entity_tracking": 17.75, "comps": 49.46, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s66f4e4e5536aa7" ] }, { "age": 60, "six": 46.1625, "five": 47.982, "tasks": { "blimp_filtered": 67.76, "supplement_filtered": 51.44, "ewok_filtered": 50.34, "entity_tracking": 19.88, "comps": 50.49, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 45.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 70, "six": 45.990833333333335, "five": 47.482, "tasks": { "blimp_filtered": 65.77, "supplement_filtered": 51.58, "ewok_filtered": 50.25, "entity_tracking": 19.29, "comps": 50.52, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 46.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 80, "six": 45.31, "five": 47.656, "tasks": { "blimp_filtered": 66.25, "supplement_filtered": 51.22, "ewok_filtered": 50.88, "entity_tracking": 18.8, "comps": 51.13, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 39.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 90, "six": 46.43, "five": 48.012, "tasks": { "blimp_filtered": 67.06, "supplement_filtered": 49.35, "ewok_filtered": 51.86, "entity_tracking": 21.04, "comps": 50.75, "global_piqa_parallel": 32.04, "global_piqa_nonparallel": 45.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] }, { "age": 100, "six": 46.1975, "five": 48.016, "tasks": { "blimp_filtered": 68.53, "supplement_filtered": 49.8, "ewok_filtered": 50.99, "entity_tracking": 19.84, "comps": 50.92, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s5d7fc09ecab70f" ] } ], "peak": { "five": { "score": 48.016, "ages": [ 100 ], "rank": 4 }, "six": { "score": 46.43, "ages": [ 90 ], "rank": 4 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 250, "id": "normal-joint/joint_12", "original_name": "joint_12", "title": "joint_12", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-joint", "cohort_label": "通常データ・joint探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 45.63583333333333, "five": 46.556, "tasks": { "blimp_filtered": 63.75, "supplement_filtered": 49.73, "ewok_filtered": 51.19, "entity_tracking": 18.44, "comps": 49.67, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 51.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s66f4e4e5536aa7" ] } ], "peak": { "five": { "score": 46.556, "ages": [ 20 ], "rank": 6 }, "six": { "score": 45.63583333333333, "ages": [ 20 ], "rank": 5 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 251, "id": "normal-joint/joint_03", "original_name": "joint_03", "title": "joint_03", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-joint", "cohort_label": "通常データ・joint探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 45.415, "five": 46.582, "tasks": { "blimp_filtered": 63.21, "supplement_filtered": 51.32, "ewok_filtered": 49.7, "entity_tracking": 18.54, "comps": 50.14, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 51.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s66f4e4e5536aa7" ] } ], "peak": { "five": { "score": 46.582, "ages": [ 20 ], "rank": 5 }, "six": { "score": 45.415, "ages": [ 20 ], "rank": 6 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 252, "id": "normal-joint/joint_06", "original_name": "joint_06", "title": "joint_06", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-joint", "cohort_label": "通常データ・joint探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 45.244166666666665, "five": 46.086, "tasks": { "blimp_filtered": 62.55, "supplement_filtered": 49.35, "ewok_filtered": 51.77, "entity_tracking": 17.29, "comps": 49.47, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 51.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s66f4e4e5536aa7" ] } ], "peak": { "five": { "score": 46.086, "ages": [ 20 ], "rank": 11 }, "six": { "score": 45.244166666666665, "ages": [ 20 ], "rank": 7 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 253, "id": "normal-joint/joint_07", "original_name": "joint_07", "title": "joint_07", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-joint", "cohort_label": "通常データ・joint探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 45.20333333333333, "five": 46.128, "tasks": { "blimp_filtered": 63.07, "supplement_filtered": 50.15, "ewok_filtered": 50.02, "entity_tracking": 17.59, "comps": 49.81, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 53.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s66f4e4e5536aa7" ] } ], "peak": { "five": { "score": 46.128, "ages": [ 20 ], "rank": 10 }, "six": { "score": 45.20333333333333, "ages": [ 20 ], "rank": 8 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 254, "id": "normal-joint/joint_01", "original_name": "joint_01", "title": "joint_01", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-joint", "cohort_label": "通常データ・joint探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 44.97083333333333, "five": 46.158, "tasks": { "blimp_filtered": 62.78, "supplement_filtered": 49.63, "ewok_filtered": 50.53, "entity_tracking": 17.68, "comps": 50.17, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s66f4e4e5536aa7" ] } ], "peak": { "five": { "score": 46.158, "ages": [ 20 ], "rank": 9 }, "six": { "score": 44.97083333333333, "ages": [ 20 ], "rank": 9 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 255, "id": "normal-joint/joint_11", "original_name": "joint_11", "title": "joint_11", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-joint", "cohort_label": "通常データ・joint探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 44.915, "five": 46.188, "tasks": { "blimp_filtered": 63.05, "supplement_filtered": 48.45, "ewok_filtered": 51.49, "entity_tracking": 18.18, "comps": 49.77, "global_piqa_parallel": 30.1, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s66f4e4e5536aa7" ] } ], "peak": { "five": { "score": 46.188, "ages": [ 20 ], "rank": 8 }, "six": { "score": 44.915, "ages": [ 20 ], "rank": 10 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 256, "id": "normal-joint/joint_05", "original_name": "joint_05", "title": "joint_05", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-joint", "cohort_label": "通常データ・joint探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 44.876666666666665, "five": 46.434, "tasks": { "blimp_filtered": 63.76, "supplement_filtered": 48.85, "ewok_filtered": 51.46, "entity_tracking": 18.11, "comps": 49.99, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 47.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s66f4e4e5536aa7" ] } ], "peak": { "five": { "score": 46.434, "ages": [ 20 ], "rank": 7 }, "six": { "score": 44.876666666666665, "ages": [ 20 ], "rank": 11 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 257, "id": "normal-joint/joint_02", "original_name": "joint_02", "title": "joint_02", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-joint", "cohort_label": "通常データ・joint探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 20, "six": 44.656666666666666, "five": 45.972, "tasks": { "blimp_filtered": 62.27, "supplement_filtered": 48.85, "ewok_filtered": 49.76, "entity_tracking": 19.08, "comps": 49.9, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 48.0 }, "checkpoint": "", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "s66f4e4e5536aa7" ] } ], "peak": { "five": { "score": 45.972, "ages": [ 20 ], "rank": 12 }, "six": { "score": 44.656666666666666, "ages": [ 20 ], "rank": 12 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 258, "id": "normal-reader/VISTARC_READER_NORMAL_T09_BF16_100M", "original_name": "VISTARC_READER_NORMAL_T09_BF16_100M", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "normal-reader", "cohort_label": "通常データ・Reader", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": false, "reader_width": 32 }, "evaluation": { "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "local_full_manifest_sha256": "6c2049fe4f27ff51a0d03ff08bdd4a63c0399ad6349220f676e25197b9dcc061", "benchmark_words": [ 100000000 ], "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "precision": "FP32, TF32 disabled; original tolerances retained; same normal-tokenizer prepared questions and pinned ranker" }, "data": { "source_dataset": "English_ARO97pct-StateReasoning3pct__800Mwords__VISTA-vocab16384-sha8e43ff69", "dataset_root": "source:s010db2cbac871a", "packed_manifest_sha256": "798b487c48f44410403d3c3081766d256aa2e29c0a36933795597d801932ebe2", "words_per_pass": 100000000, "rows": 376918, "documents": 206387, "sequence_length": 512, "prediction_targets": 154524499, "nonpad_tokens": 154956140, "document_order": "unchanged frozen first100M words; no replay or Baby grouping", "segment_id_policy": "positive stored segment IDs;0 is padding; preserve original mask exactly" }, "tokenizer": { "path": "data/tokenizer.json", "tokenizer_json_sha256": "8e43ff69bec5c685145a9f02be3d17e4eeeca3e11947ee8a15364d83427daf94", "vocab_size": 16384, "pad_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "unk_token_id": 3 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": false, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 36, "sources": [ "s9e81e276e54a41", "s237dcb71227e64", "sb577d65c0fd7c1", "se94040a1213fd5" ], "points": [ { "age": 100, "six": 46.468333333333334, "five": 48.146, "tasks": { "blimp_filtered": 67.68, "supplement_filtered": 50.07, "ewok_filtered": 52.64, "entity_tracking": 19.15, "comps": 51.19, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 48.0 }, "checkpoint": "02c8644ef7edd02292b6a940254a5c208b551c6ef9a2afa7d402b9530c2686b5", "precision": "FP32 evaluation; FP32 training through31.435341M then BF16 autocast with FP32 masters", "verification": "primary", "sources": [ "s6b6799f45fb3a8", "s01351f52c741fd", "sc97260946b0333", "s011ac913ef8ba8", "scd6d60185c8634", "sd1fd041e5f7a61", "scdb6b1aa797b13", "s06dc405264decc" ] } ], "peak": { "five": { "score": 48.146, "ages": [ 100 ], "rank": 1 }, "six": { "score": 46.468333333333334, "ages": [ 100 ], "rank": 1 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 259, "id": "normal-reader/VISTARC_READER_NORMAL_T09_100M", "original_name": "VISTARC_READER_NORMAL_T09_100M", "title": "256幅・FFN896・7層", "topology": "未確認", "recipe": "契約設定（詳細）", "cohort": "normal-reader", "cohort_label": "通常データ・Reader", "seed": 20260907, "width": 256, "ffn": 896, "depth": 7, "parts": { "width": 256, "ffn": 896, "depth": 7, "attention": "8×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 32, "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": 3.5 }, "main_lr": [], "state": "SHADOW_TEST_COMPLETE", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 256, "layers": 7, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": false, "reader_width": 32 }, "evaluation": { "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "local_full_manifest_sha256": "6c2049fe4f27ff51a0d03ff08bdd4a63c0399ad6349220f676e25197b9dcc061", "benchmark_words": [ 100000000 ], "runtime_selection": { "graph_chunk": 4, "max_batch": 256, "kv_token_budget": 32768, "prefix_graph": true }, "precision": "FP32, TF32 disabled; original tolerances retained; same normal-tokenizer prepared questions and pinned ranker" }, "data": { "source_dataset": "English_ARO97pct-StateReasoning3pct__800Mwords__VISTA-vocab16384-sha8e43ff69", "dataset_root": "source:s010db2cbac871a", "packed_manifest_sha256": "798b487c48f44410403d3c3081766d256aa2e29c0a36933795597d801932ebe2", "words_per_pass": 100000000, "rows": 376918, "documents": 206387, "sequence_length": 512, "prediction_targets": 154524499, "nonpad_tokens": 154956140, "document_order": "unchanged frozen first100M words; no replay or Baby grouping", "segment_id_policy": "positive stored segment IDs;0 is padding; preserve original mask exactly" }, "tokenizer": { "path": "data/tokenizer.json", "tokenizer_json_sha256": "8e43ff69bec5c685145a9f02be3d17e4eeeca3e11947ee8a15364d83427daf94", "vocab_size": 16384, "pad_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "unk_token_id": 3 }, "runtime": { "training_precision": "fp32", "evaluation_precision": "fp32", "tf32": false, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false }, "condition_conflicts": [], "profile": 36, "sources": [ "sf551f0b3d0d6d4", "sbc6a84b2e8f25d", "s2914ceebaddc16", "s961a27a7ab498d", "s46174b954ac0ba", "s5c33ec1b9a8d1e", "s04c342d627bec3" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 260, "id": "normal-threshold/threshold_09", "original_name": "threshold_09", "title": "threshold_09", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-threshold", "cohort_label": "通常データ・閾値探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 100, "six": 47.24166666666667, "five": 48.972, "tasks": { "blimp_filtered": 68.05, "supplement_filtered": 50.44, "ewok_filtered": 53.59, "entity_tracking": 21.3, "comps": 51.48, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 50.0 }, "checkpoint": "162e407127843cf3b1d334ff9e31bbe95f7c4e2c562531414fead52a35fd7e60", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sd6676f25136e60" ] } ], "peak": { "five": { "score": 48.972, "ages": [ 100 ], "rank": 1 }, "six": { "score": 47.24166666666667, "ages": [ 100 ], "rank": 1 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 261, "id": "normal-threshold/top3_fixed_alpha01", "original_name": "top3_fixed_alpha01", "title": "top3_fixed_alpha01", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-threshold", "cohort_label": "通常データ・閾値探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 100, "six": 46.94083333333333, "five": 48.322, "tasks": { "blimp_filtered": 67.46, "supplement_filtered": 50.62, "ewok_filtered": 51.59, "entity_tracking": 20.39, "comps": 51.55, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 49.0 }, "checkpoint": "758dda2f8ab454e558d129591f2d1d4475ee95b31ebdfb0da2b84310990764aa", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sd6676f25136e60" ] } ], "peak": { "five": { "score": 48.322, "ages": [ 100 ], "rank": 5 }, "six": { "score": 46.94083333333333, "ages": [ 100 ], "rank": 2 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 262, "id": "normal-threshold/threshold_01", "original_name": "threshold_01", "title": "threshold_01", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-threshold", "cohort_label": "通常データ・閾値探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 100, "six": 46.92166666666667, "five": 48.482, "tasks": { "blimp_filtered": 68.02, "supplement_filtered": 51.39, "ewok_filtered": 51.84, "entity_tracking": 20.03, "comps": 51.13, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 53.0 }, "checkpoint": "46da3d186a9bae084d3cc8c7a995841df43f9801c10335cd47b3296008b6fc4b", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sed3208ffa4b155" ] } ], "peak": { "five": { "score": 48.482, "ages": [ 100 ], "rank": 4 }, "six": { "score": 46.92166666666667, "ages": [ 100 ], "rank": 3 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 263, "id": "normal-threshold/threshold_fixed_alpha01", "original_name": "threshold_fixed_alpha01", "title": "threshold_fixed_alpha01", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-threshold", "cohort_label": "通常データ・閾値探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 100, "six": 46.630833333333335, "five": 48.53, "tasks": { "blimp_filtered": 67.81, "supplement_filtered": 51.66, "ewok_filtered": 51.72, "entity_tracking": 19.98, "comps": 51.48, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 50.0 }, "checkpoint": "ad6684dfc949a1a075cfbf9f3212364e4c81f00846eec417262054b839a9bc5c", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sd6676f25136e60" ] } ], "peak": { "five": { "score": 48.53, "ages": [ 100 ], "rank": 3 }, "six": { "score": 46.630833333333335, "ages": [ 100 ], "rank": 4 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 264, "id": "normal-threshold/threshold_02", "original_name": "threshold_02", "title": "threshold_02", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-threshold", "cohort_label": "通常データ・閾値探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 100, "six": 46.464166666666664, "five": 48.63, "tasks": { "blimp_filtered": 67.6, "supplement_filtered": 50.98, "ewok_filtered": 52.58, "entity_tracking": 20.81, "comps": 51.18, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 47.0 }, "checkpoint": "fee462c84b5660c91df2fd3b9b4b4f68f17b8c8853b1c4a1f9cd5d6c7d378a2a", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sed3208ffa4b155" ] } ], "peak": { "five": { "score": 48.63, "ages": [ 100 ], "rank": 2 }, "six": { "score": 46.464166666666664, "ages": [ 100 ], "rank": 5 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 265, "id": "normal-threshold/full_fixed_alpha01", "original_name": "full_fixed_alpha01", "title": "full_fixed_alpha01", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-threshold", "cohort_label": "通常データ・閾値探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 100, "six": 46.443333333333335, "five": 48.316, "tasks": { "blimp_filtered": 68.39, "supplement_filtered": 49.52, "ewok_filtered": 52.41, "entity_tracking": 20.03, "comps": 51.23, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 46.0 }, "checkpoint": "932c92e190d59c5419fd65903d23250415b5ee49da99749b5ad4186bc1ccae2e", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sd6676f25136e60" ] } ], "peak": { "five": { "score": 48.316, "ages": [ 100 ], "rank": 6 }, "six": { "score": 46.443333333333335, "ages": [ 100 ], "rank": 6 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 266, "id": "normal-threshold/threshold_08", "original_name": "threshold_08", "title": "threshold_08", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-threshold", "cohort_label": "通常データ・閾値探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 100, "six": 46.196666666666665, "five": 48.112, "tasks": { "blimp_filtered": 67.81, "supplement_filtered": 48.32, "ewok_filtered": 53.62, "entity_tracking": 19.55, "comps": 51.26, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 48.0 }, "checkpoint": "eae9fcea09c8e90d4a19e132ff9ca7135c49758a92021dc7049e33a35b3a65cb", "precision": null, "verification": "歴史台帳・原source併記", "sources": [ "s6b6799f45fb3a8", "sd6676f25136e60" ] } ], "peak": { "five": { "score": 48.112, "ages": [ 100 ], "rank": 7 }, "six": { "score": 46.196666666666665, "ages": [ 100 ], "rank": 7 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 267, "id": "normal-threshold/standard_vista", "original_name": "standard_vista", "title": "standard_vista", "topology": "未確認", "recipe": "設定の元資料を参照", "cohort": "normal-threshold", "cohort_label": "通常データ・閾値探索", "seed": null, "width": "未確認", "ffn": "未確認", "depth": "未確認", "parts": { "width": "未確認", "ffn": "未確認", "depth": "未確認", "attention": "未確認×未確認", "router": "未確認", "candidate": "未確認", "reader": "未確認", "received": "未確認", "output": "未確認", "reference": "未確認", "ffn_add": "未確認", "qk": "未確認", "distance": "未確認", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "未確認", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "未確認", "ffn_ratio": null }, "main_lr": [], "state": "採点済み", "architecture": {}, "evaluation": null, "data": null, "tokenizer": null, "runtime": null, "condition_conflicts": [], "profile": 35, "sources": [], "points": [ { "age": 100, "six": 45.86, "five": 47.416, "tasks": { "blimp_filtered": 67.46, "supplement_filtered": 48.74, "ewok_filtered": 50.56, "entity_tracking": 19.31, "comps": 51.01, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 48.0 }, "checkpoint": "95754dc3ee7da655e1c035b7a33445226aa7926b4df16ca61e859da06ecd9c7f", "precision": null, "verification": "primary", "sources": [ "s6b6799f45fb3a8", "sd6676f25136e60", "s52fb3c721aa31a" ] } ], "peak": { "five": { "score": 47.416, "ages": [ 100 ], "rank": 8 }, "six": { "score": 45.86, "ages": [ 100 ], "rank": 8 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 268, "id": "run/IPRLM384_B_EIGHT_OTHER_RESET0_SEED20260907_80M_20261025", "original_name": "IPRLM384_B_EIGHT_OTHER_RESET0_SEED20260907_80M_20261025", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要−10％・その他＋5％／指定8項目だけ0％", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M,70M,80M after training" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000, 70000000, 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000, 70000000, 80000000 ], "timing": "after80M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Parent#3 matched ages, baseline#31, allknownmodel inventory; observedpeak and taskgroups distinct" }, "milestones_words": [ 40000000, 70000000, 80000000 ], "checkpoints_words": [ 40000000, 70000000, 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 99, "sources": [ "s7704f167d2671e" ], "points": [ { "age": 40.0, "six": 48.115833333333335, "five": 50.812, "tasks": { "blimp_filtered": 68.79, "supplement_filtered": 58.66, "ewok_filtered": 52.97, "entity_tracking": 21.4, "comps": 52.24, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 45.0 }, "checkpoint": "742f90a737107e4e6622264f1903b9814772953185f6b77dc18b7e4e3c76fd9d", "precision": null, "verification": "primary", "sources": [ "s82c498f1831d8a" ] }, { "age": 70.0, "six": 48.5625, "five": 51.048, "tasks": { "blimp_filtered": 69.75, "supplement_filtered": 61.27, "ewok_filtered": 50.88, "entity_tracking": 20.57, "comps": 52.77, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 48.0 }, "checkpoint": "07053e114c370ce7ac5e0d3847662a439babf2ae697264b1283f947811795d3a", "precision": null, "verification": "primary", "sources": [ "s82c498f1831d8a" ] }, { "age": 80.0, "six": 48.946666666666665, "five": 51.306, "tasks": { "blimp_filtered": 69.41, "supplement_filtered": 61.9, "ewok_filtered": 52.81, "entity_tracking": 19.91, "comps": 52.5, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "5c49985b32e6661baca26455a92dacb199ea316e23eee0c4d1cad7075300f23a", "precision": null, "verification": "primary", "sources": [ "s82c498f1831d8a" ] } ], "peak": { "five": { "score": 51.306, "ages": [ 80.0 ], "rank": 19 }, "six": { "score": 48.946666666666665, "ages": [ 80.0 ], "rank": 40 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 269, "id": "run/IPRLM384_B_FIVE_RESET0_ALPHA_TAU_ATTNBIAS105_SEED20260907_80M_20261025", "original_name": "IPRLM384_B_FIVE_RESET0_ALPHA_TAU_ATTNBIAS105_SEED20260907_80M_20261025", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要−10％・その他＋5％／5項目0％・α/τ/Attention gate bias＋5％", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at40M,70M,80M after training" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 40000000, 70000000, 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 40000000, 70000000, 80000000 ], "timing": "after80M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "Parent#3 matched ages, baseline#31, allknownmodel inventory; observedpeak and taskgroups distinct" }, "milestones_words": [ 40000000, 70000000, 80000000 ], "checkpoints_words": [ 40000000, 70000000, 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 101, "sources": [ "s6354c49d09e7ef" ], "points": [ { "age": 70.0, "six": 48.60916666666667, "five": 51.298, "tasks": { "blimp_filtered": 69.67, "supplement_filtered": 60.13, "ewok_filtered": 52.16, "entity_tracking": 22.08, "comps": 52.45, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 48.0 }, "checkpoint": "4404ee686ed88e8d352bbfc14ab5a1fd1c876db1f03c5ee5970b4325a5d14509", "precision": null, "verification": "primary", "sources": [ "sddc19160c03f9d" ] }, { "age": 80.0, "six": 48.5075, "five": 51.076, "tasks": { "blimp_filtered": 68.59, "supplement_filtered": 60.82, "ewok_filtered": 52.44, "entity_tracking": 21.12, "comps": 52.41, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 49.0 }, "checkpoint": "d30245cdc30c20a048bce266d0ba6b7073d10e7fb3905819bc43ff442d1f9c12", "precision": null, "verification": "primary", "sources": [ "sddc19160c03f9d" ] } ], "peak": { "five": { "score": 51.298, "ages": [ 70.0 ], "rank": 20 }, "six": { "score": 48.60916666666667, "ages": [ 70.0 ], "rank": 68 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 270, "id": "run/MODEL3_P_SMALL_100M_20261025", "original_name": "MODEL3_P_SMALL_100M_20261025", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "未確認", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "未確認", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 117, "plastic_rank": 16, "plastic_state": "owner-local F", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input" }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "a8280b292806610a2e7d5e3c9d21dd42583776bda888159938666b5202177d25", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source_v2", "GPU": "RTX3090", "max_vram_gib": 22 }, "condition_conflicts": [], "profile": 126, "sources": [ "s0bc57808dce1f6", "sb45fd3801ddee6" ], "points": [ { "age": 70.0, "six": 48.98083333333334, "five": 51.644, "tasks": { "blimp_filtered": 70.01, "supplement_filtered": 60.78, "ewok_filtered": 54.41, "entity_tracking": 20.32, "comps": 52.7, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 49.0 }, "checkpoint": "4056b3372b00ee930b12cca8a1cd3aa9c4035371e33e110a3900f03d76157787", "precision": "FP32", "verification": "primary", "sources": [ "s2ca6db766194c0", "s64c54e26344d82" ] }, { "age": 80.0, "six": 48.454166666666666, "five": 51.006, "tasks": { "blimp_filtered": 68.52, "supplement_filtered": 61.63, "ewok_filtered": 53.16, "entity_tracking": 19.07, "comps": 52.65, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 51.0 }, "checkpoint": "0a64a9030bcd27169fea00866f8e588089e16043aac9b0ae2a1fa853025f335e", "precision": "FP32", "verification": "primary", "sources": [ "s58c04c065217ac", "s07df67dc6998d1" ] }, { "age": 90.0, "six": 48.41416666666667, "five": 50.564, "tasks": { "blimp_filtered": 69.15, "supplement_filtered": 59.29, "ewok_filtered": 51.93, "entity_tracking": 20.54, "comps": 51.91, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 53.0 }, "checkpoint": "fe7e0c7139d22b58641071f10a1dd0bc56cd4e216b9f90f9935c15f45f2e22af", "precision": "FP32", "verification": "primary", "sources": [ "s6e3a9652855cde", "s648e5da66fa64b" ] }, { "age": 100.0, "six": 48.681666666666665, "five": 50.894, "tasks": { "blimp_filtered": 69.27, "supplement_filtered": 60.26, "ewok_filtered": 52.11, "entity_tracking": 20.27, "comps": 52.56, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 50.0 }, "checkpoint": "66964c020938d31a8c2f9854b42129a13034ae14f2c8774dc1437cac26e838ff", "precision": "FP32", "verification": "primary", "sources": [ "sf915f0fc8eac4f", "s463abac900adeb" ] } ], "peak": { "five": { "score": 51.644, "ages": [ 70.0 ], "rank": 5 }, "six": { "score": 48.98083333333334, "ages": [ 70.0 ], "rank": 37 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 271, "id": "run/IPRLM384_B_FFN_SPREAD_PLUS5_FROM269_SEED20260907_80M_20261025", "original_name": "IPRLM384_B_FFN_SPREAD_PLUS5_FROM269_SEED20260907_80M_20261025", "title": "384幅・FFN1408・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 9, "parts": { "width": 384, "ffn": 1408, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "評価記録なし", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_scaling_anchor_hidden": 1120, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes at70M,80M after training" ], "intermediate": "none", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000 ], "timing": "after80M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "matched #269 and complete all-model inventory; observed peak and per-task groups distinct" }, "milestones_words": [ 70000000, 80000000 ], "checkpoints_words": [ 70000000, 80000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 100, "sources": [ "s65679e44fa5e9c" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 272, "id": "run/IPRLM384_MODEL3_DEPTH10_FFN1408_SEED20260907_100M_20261027", "original_name": "IPRLM384_MODEL3_DEPTH10_FFN1408_SEED20260907_100M_20261027", "title": "384幅・FFN1408・10層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 10, "parts": { "width": 384, "ffn": 1408, "depth": 10, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "10層 / 9・10層 / 8・9・10層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 10 ], [ 9, 10 ], [ 8, 9, 10 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 10, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 10 ], [ 9, 10 ], [ 8, 9, 10 ] ], "ffn_scaling_anchor_hidden": 1120, "initialization_depth_anchor": 9, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local6 at70/80/90/100M", "supervised7 at highest-six checkpoint" ], "intermediate": "none; score after100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "full known inventory with #3 parent; seven-task evaluated cohort also compared" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ], "checkpoints_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 130, "sources": [ "s9df3f61d3b5c3c" ], "points": [ { "age": 70.0, "six": 48.594166666666666, "five": 51.18, "tasks": { "blimp_filtered": 70.52, "supplement_filtered": 61.28, "ewok_filtered": 51.78, "entity_tracking": 19.9, "comps": 52.42, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 49.0 }, "checkpoint": "51ca8f73acede3c31ea8cd65453e3d1c01c6afdc95953d8243d491a432782834", "precision": null, "verification": "primary", "sources": [ "s159a0455fd0ea7" ] }, { "age": 80.0, "six": 49.01583333333333, "five": 51.48, "tasks": { "blimp_filtered": 69.31, "supplement_filtered": 63.56, "ewok_filtered": 51.52, "entity_tracking": 20.19, "comps": 52.82, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 53.0 }, "checkpoint": "43b175d1028bf4351bbe1daa8280dbc1357ed5f4acba7eb6d28fc9059d818229", "precision": null, "verification": "primary", "sources": [ "s159a0455fd0ea7" ] }, { "age": 90.0, "six": 46.891666666666666, "five": 49.834, "tasks": { "blimp_filtered": 67.97, "supplement_filtered": 57.15, "ewok_filtered": 51.99, "entity_tracking": 20.0, "comps": 52.06, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 43.0 }, "checkpoint": "0f7244893432e46b75942a4a6b72d2fecd205a47eaa8f6031eb4d30b3c3600ba", "precision": null, "verification": "primary", "sources": [ "s159a0455fd0ea7" ] }, { "age": 100.0, "six": 48.92583333333334, "five": 51.278, "tasks": { "blimp_filtered": 68.69, "supplement_filtered": 61.69, "ewok_filtered": 52.64, "entity_tracking": 20.81, "comps": 52.56, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "checkpoint": "321f48cc1079a33c68a11a04191dbe67f2f4c34507e7aa65972ae5a1f41c869d", "precision": null, "verification": "primary", "sources": [ "s159a0455fd0ea7" ] } ], "peak": { "five": { "score": 51.48, "ages": [ 80.0 ], "rank": 9 }, "six": { "score": 49.01583333333333, "ages": [ 80.0 ], "rank": 32 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 273, "id": "run/IPRLM384_MODEL3_DEPTH11_FFN1408_SEED20260907_100M_20261027", "original_name": "IPRLM384_MODEL3_DEPTH11_FFN1408_SEED20260907_100M_20261027", "title": "384幅・FFN1408・11層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 11, "parts": { "width": 384, "ffn": 1408, "depth": 11, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "11層 / 10・11層 / 9・10・11層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 11 ], [ 10, 11 ], [ 9, 10, 11 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 11, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 11 ], [ 10, 11 ], [ 9, 10, 11 ] ], "ffn_scaling_anchor_hidden": 1120, "initialization_depth_anchor": 9, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local6 at70/80/90/100M" ], "intermediate": "none; score after100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "all known models, primary parent#3 and depth10#272; same seed; no supervised7" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ], "checkpoints_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 131, "sources": [ "s36b2fdfe9c4c11" ], "points": [ { "age": 70.0, "six": 47.72, "five": 50.722, "tasks": { "blimp_filtered": 70.87, "supplement_filtered": 58.59, "ewok_filtered": 51.29, "entity_tracking": 20.56, "comps": 52.3, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 46.0 }, "checkpoint": "c8746d500f56b5aead2d08080e2fb213d0bd2bd81e8043696dce5dd0d1835c2b", "precision": null, "verification": "primary", "sources": [ "s14464eb47ccfab" ] }, { "age": 80.0, "six": 49.115, "five": 50.922, "tasks": { "blimp_filtered": 69.08, "supplement_filtered": 60.62, "ewok_filtered": 52.66, "entity_tracking": 19.85, "comps": 52.4, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 52.0 }, "checkpoint": "88cd647734e13cd1a8b1e04ef78847e469b5066e8b8efe08258df6dedfc2c08c", "precision": null, "verification": "primary", "sources": [ "s14464eb47ccfab" ] }, { "age": 90.0, "six": 47.94166666666667, "five": 50.394, "tasks": { "blimp_filtered": 69.88, "supplement_filtered": 56.8, "ewok_filtered": 52.07, "entity_tracking": 20.84, "comps": 52.38, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 50.0 }, "checkpoint": "f5229d4f11eaebcd5ad77c014664a260872341b46fef45bd5de573eecb04aa63", "precision": null, "verification": "primary", "sources": [ "s14464eb47ccfab" ] }, { "age": 100.0, "six": 48.19833333333333, "five": 50.008, "tasks": { "blimp_filtered": 68.86, "supplement_filtered": 58.42, "ewok_filtered": 52.09, "entity_tracking": 18.96, "comps": 51.71, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 55.0 }, "checkpoint": "20adce22a095e4b3708126316cda8f6dfce1094aec4ddbaac4a095c700bdffc0", "precision": null, "verification": "primary", "sources": [ "s14464eb47ccfab" ] } ], "peak": { "five": { "score": 50.922, "ages": [ 80.0 ], "rank": 55 }, "six": { "score": 49.115, "ages": [ 80.0 ], "rank": 24 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 274, "id": "run/IPRLM384_MODEL3_DEPTH12_FFN1408_SEED20260907_100M_20261027", "original_name": "IPRLM384_MODEL3_DEPTH12_FFN1408_SEED20260907_100M_20261027", "title": "384幅・FFN1408・12層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 12, "parts": { "width": 384, "ffn": 1408, "depth": 12, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "12層 / 11・12層 / 10・11・12層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 12 ], [ 11, 12 ], [ 10, 11, 12 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 12, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 12 ], [ 11, 12 ], [ 10, 11, 12 ] ], "ffn_scaling_anchor_hidden": 1120, "initialization_depth_anchor": 9, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local6 at70/80/90/100M" ], "intermediate": "none; score after100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All known models; parent#3,depth10#272,depth11#273; matched seed; no supervised7" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ], "checkpoints_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 132, "sources": [ "s5b186c0ef80216" ], "points": [ { "age": 70.0, "six": 48.2225, "five": 50.634, "tasks": { "blimp_filtered": 70.61, "supplement_filtered": 59.11, "ewok_filtered": 50.91, "entity_tracking": 19.78, "comps": 52.76, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 50.0 }, "checkpoint": "167d83c08c2017ab2f751d33c57d29154170a708f925c6c522e841d8b4871bd9", "precision": null, "verification": "primary", "sources": [ "s2a6067e4cee7e0" ] }, { "age": 80.0, "six": 48.82333333333333, "five": 51.258, "tasks": { "blimp_filtered": 69.99, "supplement_filtered": 61.21, "ewok_filtered": 51.33, "entity_tracking": 21.37, "comps": 52.39, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "9542a47c4eeb27e1d63b25a3cde7f995e9e4b428012fc6e033928727a72aac04", "precision": null, "verification": "primary", "sources": [ "s2a6067e4cee7e0" ] }, { "age": 90.0, "six": 47.61333333333334, "five": 50.288, "tasks": { "blimp_filtered": 70.21, "supplement_filtered": 57.31, "ewok_filtered": 52.08, "entity_tracking": 19.48, "comps": 52.36, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 51.0 }, "checkpoint": "b321d75ce4e27b89887f625639c29bfa25e3078bde76eb21348900b1c71e70c0", "precision": null, "verification": "primary", "sources": [ "s2a6067e4cee7e0" ] }, { "age": 100.0, "six": 48.575, "five": 50.866, "tasks": { "blimp_filtered": 69.54, "supplement_filtered": 60.19, "ewok_filtered": 51.47, "entity_tracking": 20.74, "comps": 52.39, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 49.0 }, "checkpoint": "81c779cf4170739f43ce97cc48a6589a9b8cc95893cf4bf62725154eecc798e0", "precision": null, "verification": "primary", "sources": [ "s2a6067e4cee7e0" ] } ], "peak": { "five": { "score": 51.258, "ages": [ 80.0 ], "rank": 22 }, "six": { "score": 48.82333333333333, "ages": [ 80.0 ], "rank": 51 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 275, "id": "run/IPRLM384_MODEL3_DEPTH8_FFN1408_SEED20260907_100M_20261027", "original_name": "IPRLM384_MODEL3_DEPTH8_FFN1408_SEED20260907_100M_20261027", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "initialization_depth_anchor": 9, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local6 at70/80/90/100M" ], "intermediate": "none; score after100M", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "no_endpoint_penalty": true, "parent_peak_five_axis": 52.002, "same_family_peak_six_axis": 49.8675, "comparison_scope": "All known models, parent#3 and depth10#272/depth11#273/depth12#274; same seed, no supervised7" }, "milestones_words": [ 70000000, 80000000, 90000000, 100000000 ], "checkpoints_words": [ 70000000, 80000000, 90000000, 100000000 ] }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260 }, "condition_conflicts": [], "profile": 133, "sources": [ "sc1fa03494cdd60" ], "points": [ { "age": 70.0, "six": 47.538333333333334, "five": 50.222, "tasks": { "blimp_filtered": 69.27, "supplement_filtered": 58.63, "ewok_filtered": 50.98, "entity_tracking": 19.96, "comps": 52.27, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 43.0 }, "checkpoint": "d0d6b90a1b0234d4e20a884403f4ed0972561baa7977e159427b9358cd362283", "precision": null, "verification": "primary", "sources": [ "sca61306e756591" ] }, { "age": 80.0, "six": 48.380833333333335, "five": 50.624, "tasks": { "blimp_filtered": 68.47, "supplement_filtered": 59.48, "ewok_filtered": 52.27, "entity_tracking": 20.44, "comps": 52.46, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "checkpoint": "7eae35696055fa94b9c2eb7b26b67a2563ffe8c386b5fedc65922ec49f62c3c4", "precision": null, "verification": "primary", "sources": [ "sca61306e756591" ] }, { "age": 90.0, "six": 48.13166666666667, "five": 50.34, "tasks": { "blimp_filtered": 69.63, "supplement_filtered": 57.23, "ewok_filtered": 52.07, "entity_tracking": 20.89, "comps": 51.88, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 47.0 }, "checkpoint": "c3ee35a19ef9d4551cb8f96bdcf2d64cde5b30449ce3e3e6f1598dc56b117d4d", "precision": null, "verification": "primary", "sources": [ "sca61306e756591" ] }, { "age": 100.0, "six": 48.05, "five": 50.042, "tasks": { "blimp_filtered": 69.38, "supplement_filtered": 56.76, "ewok_filtered": 51.52, "entity_tracking": 20.04, "comps": 52.51, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 49.0 }, "checkpoint": "9c2f058ca48c837ea05bbccba5761a36289cd29a775cc3eb91754da69061e327", "precision": null, "verification": "primary", "sources": [ "sca61306e756591" ] } ], "peak": { "five": { "score": 50.624, "ages": [ 80.0 ], "rank": 93 }, "six": { "score": 48.380833333333335, "ages": [ 80.0 ], "rank": 86 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 276, "id": "run/IPRLM384_MODEL270_DEPTH10_FFN1408_SEED20260907_100M_20261027", "original_name": "IPRLM384_MODEL270_DEPTH10_FFN1408_SEED20260907_100M_20261027", "title": "384幅・FFN1408・10層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 10, "parts": { "width": 384, "ffn": 1408, "depth": 10, "attention": "12×32", "router": 32, "candidate": "未確認", "reader": 48, "received": "あり", "output": "MoS3", "reference": "10層 / 9・10層 / 8・9・10層", "ffn_add": "未確認", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 10 ], [ 9, 10 ], [ 8, 9, 10 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 10, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 10 ], [ 9, 10 ], [ 8, 9, 10 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 130, "plastic_rank": 16, "plastic_state": "owner-local F", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9 }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "a8280b292806610a2e7d5e3c9d21dd42583776bda888159938666b5202177d25", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "all known models and#270,#3,#272-275", "supervised7": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source_v2", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32 }, "condition_conflicts": [], "profile": 125, "sources": [ "s6f8abac56a8e5e" ], "points": [ { "age": 70.0, "six": 48.21, "five": 50.422, "tasks": { "blimp_filtered": 69.44, "supplement_filtered": 59.59, "ewok_filtered": 51.57, "entity_tracking": 19.5, "comps": 52.01, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "98c4dd77e371ef4acd07a45e920d3827a00e5c0770032c7bed45f36719702ec1", "precision": "FP32", "verification": "primary", "sources": [ "s89079d462e6dd2", "sbaeb431918986a" ] }, { "age": 80.0, "six": 48.28333333333333, "five": 50.698, "tasks": { "blimp_filtered": 67.31, "supplement_filtered": 62.62, "ewok_filtered": 51.45, "entity_tracking": 19.64, "comps": 52.47, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 53.0 }, "checkpoint": "a736e900e9f67c2be1e920aba65ef5c33e8695e18b3f8540282fedbb31ed32d5", "precision": "FP32", "verification": "primary", "sources": [ "s7f3f8e790314a7", "sbaeb431918986a" ] }, { "age": 90.0, "six": 47.215833333333336, "five": 49.62, "tasks": { "blimp_filtered": 67.64, "supplement_filtered": 59.04, "ewok_filtered": 49.98, "entity_tracking": 19.54, "comps": 51.9, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 50.0 }, "checkpoint": "8cf717877d2dd991dd27f0105c9ab7bcc84856f50b773efc6938c39cab92dd99", "precision": "FP32", "verification": "primary", "sources": [ "s7dfbb2e62c3299", "sbaeb431918986a" ] }, { "age": 100.0, "six": 48.71, "five": 50.334, "tasks": { "blimp_filtered": 67.88, "supplement_filtered": 60.02, "ewok_filtered": 52.33, "entity_tracking": 19.61, "comps": 51.83, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 54.0 }, "checkpoint": "49b83dc137da50e76df09bafafbd99d3f8933541d5a4557721006f6ff2914eba", "precision": "FP32", "verification": "primary", "sources": [ "s2cd0811b47a7e5", "sbaeb431918986a" ] } ], "peak": { "five": { "score": 50.698, "ages": [ 80.0 ], "rank": 83 }, "six": { "score": 48.71, "ages": [ 100.0 ], "rank": 62 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 277, "id": "run/IPRLM384_MODEL270_DEPTH8_FFN1408_SEED20260907_100M_20261028", "original_name": "IPRLM384_MODEL270_DEPTH8_FFN1408_SEED20260907_100M_20261028", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 104, "plastic_rank": 16, "plastic_state": "owner-local F", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9, "ffn_additive_bypass": false, "candidate_handling": "候補の平均・分散を帰還" }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "a8280b292806610a2e7d5e3c9d21dd42583776bda888159938666b5202177d25", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "all known models; parent#270,plastic10L#276,plain8L#275 and#3 depth-series", "supervised7": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source_v2", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "all+outer_cuda_graph; compiled forward/backward only, AdamW/clip and metadata validation outside capture", "cuda_graph_origin": "#276/extensions/steady_speed_20261028/graph_update.py; revalidated at8layers before fresh training" }, "condition_conflicts": [], "profile": 70, "sources": [ "se192ad303f865a", "sec69c3778b413a" ], "points": [ { "age": 70.0, "six": 49.486666666666665, "five": 51.66, "tasks": { "blimp_filtered": 69.06, "supplement_filtered": 62.15, "ewok_filtered": 52.68, "entity_tracking": 21.52, "comps": 52.89, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 52.0 }, "checkpoint": "dec13ad485462a5e010781b616b1dbb8c59fb179b16246b6a3fb37e033c96569", "precision": "FP32", "verification": "primary", "sources": [ "sa2193dc462f30b", "sca5baf5d2ad897", "s38973c75d228ce" ] }, { "age": 80.0, "six": 49.770833333333336, "five": 51.598, "tasks": { "blimp_filtered": 68.56, "supplement_filtered": 62.54, "ewok_filtered": 53.63, "entity_tracking": 20.43, "comps": 52.83, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 57.0 }, "checkpoint": "f49f86102b1646e6b067b4df261e627e2807dfe7a430e216af137bcbd408fcdc", "precision": "FP32", "verification": "primary", "sources": [ "s834942158210cd", "sca5baf5d2ad897", "s38973c75d228ce" ] }, { "age": 90.0, "six": 47.69166666666667, "five": 50.106, "tasks": { "blimp_filtered": 67.46, "supplement_filtered": 58.24, "ewok_filtered": 52.43, "entity_tracking": 20.14, "comps": 52.26, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 46.0 }, "checkpoint": "296bbde80b4d306e4ef340b36f81e0fb2cf739d105e713c7b47ec51409386903", "precision": "FP32", "verification": "primary", "sources": [ "s12d01c84bf06ea", "sca5baf5d2ad897", "s38973c75d228ce" ] }, { "age": 100.0, "six": 48.67333333333333, "five": 50.484, "tasks": { "blimp_filtered": 67.84, "supplement_filtered": 57.01, "ewok_filtered": 53.17, "entity_tracking": 21.99, "comps": 52.41, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 54.0 }, "checkpoint": "742debccba392e3065d801052033572d6efb08f44e56ca4328ae93faecdd599c", "precision": "FP32", "verification": "primary", "sources": [ "sa56b952bc82d9e", "sca5baf5d2ad897", "s38973c75d228ce" ] } ], "peak": { "five": { "score": 51.66, "ages": [ 70.0 ], "rank": 3 }, "six": { "score": 49.770833333333336, "ages": [ 80.0 ], "rank": 2 } }, "clear_points": [ { "number": 277, "age": 70, "seed": 20260907, "checkpoint_sha256": "dec13ad485462a5e010781b616b1dbb8c59fb179b16246b6a3fb37e033c96569", "old_six": 49.486666666666665, "old_five": 51.66, "old_tasks": { "blimp_filtered": 69.06, "supplement_filtered": 62.15, "ewok_filtered": 52.68, "entity_tracking": 21.52, "comps": 52.89, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 52.0 }, "correct": 12181, "total": 17274, "accuracy": 70.51638300335765, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6879, "total": 8392, "ties": 0 }, "comps": { "correct": 3891, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 0, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 15, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1017, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 69, "total": 114, "ties": 0 }, "boolq": { "correct": 9, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 238.7687734889987, "checkpoint_rank": 17 }, { "number": 277, "age": 80, "seed": 20260907, "checkpoint_sha256": "f49f86102b1646e6b067b4df261e627e2807dfe7a430e216af137bcbd408fcdc", "old_six": 49.770833333333336, "old_five": 51.598, "old_tasks": { "blimp_filtered": 68.56, "supplement_filtered": 62.54, "ewok_filtered": 53.63, "entity_tracking": 20.43, "comps": 52.83, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 57.0 }, "correct": 12651, "total": 17274, "accuracy": 73.23723515109413, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6852, "total": 8392, "ties": 0 }, "comps": { "correct": 3942, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 15, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 989, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 6, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 4, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 2, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 0, "total": 2, "ties": 0 }, "piqa1838": { "correct": 73, "total": 114, "ties": 0 }, "boolq": { "correct": 8, "total": 53, "ties": 0 }, "mnli": { "correct": 25, "total": 203, "ties": 0 }, "mrpc": { "correct": 5, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 679, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 260.8200847420012, "checkpoint_rank": 1, "rank": 1 } ], "clear_peak": { "number": 277, "age": 80, "seed": 20260907, "checkpoint_sha256": "f49f86102b1646e6b067b4df261e627e2807dfe7a430e216af137bcbd408fcdc", "old_six": 49.770833333333336, "old_five": 51.598, "old_tasks": { "blimp_filtered": 68.56, "supplement_filtered": 62.54, "ewok_filtered": 53.63, "entity_tracking": 20.43, "comps": 52.83, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 57.0 }, "correct": 12651, "total": 17274, "accuracy": 73.23723515109413, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6852, "total": 8392, "ties": 0 }, "comps": { "correct": 3942, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 15, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 989, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 6, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 4, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 2, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 0, "total": 2, "ties": 0 }, "piqa1838": { "correct": 73, "total": 114, "ties": 0 }, "boolq": { "correct": 8, "total": 53, "ties": 0 }, "mnli": { "correct": 25, "total": 203, "ties": 0 }, "mrpc": { "correct": 5, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 679, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 260.8200847420012, "checkpoint_rank": 1, "rank": 1 }, "present": true }, { "number": 278, "id": "run/IPRLM_MODEL1_STRUCTURE_LR3_MAJOR480_OTHER105_SEED20260907_100M_20261028", "original_name": "IPRLM_MODEL1_STRUCTURE_LR3_MAJOR480_OTHER105_SEED20260907_100M_20261028", "title": "320幅・FFN1120・9層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 320, "ffn": 1120, "depth": 9, "parts": { "width": 320, "ffn": 1120, "depth": 9, "attention": "10×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "9層 / 8・9層 / 7・8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "中止・削除済み（ユーザー指定）", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 320, "layers": 9, "hidden": 1120, "attention_heads": 10, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 320, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 9 ], [ 8, 9 ], [ 7, 8, 9 ] ], "ffn_additive_bypass": false, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes", "SyntaxGym31", "original PIQA1838", "frozen bAbI1947 and ProofWriter1000 at80M" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 strict recurrent prefix graph, verified for native branch123 head", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 32768, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.715, "comparison_scope": "Matched ages and observed peaks separately, including ordinary320 shared-last2 and all selected historical models" }, "milestones_words": [ 50000000, 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "comparison": "all known models and original#1,#2,#3; same ages and independent observed5/6peaks", "supervised7": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260, "recompile_limit": 32, "fallback_random": true }, "condition_conflicts": [], "profile": 147, "sources": [ "sb5d23783b52d22" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 279, "id": "run/IPRLM_MODEL2_STRUCTURE_LR3_MAJOR480_OTHER105_SEED20260907_100M_20261028", "original_name": "IPRLM_MODEL2_STRUCTURE_LR3_MAJOR480_OTHER105_SEED20260907_100M_20261028", "title": "384幅・FFN1344・9層", "topology": "MoS3・共通2層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1344, "depth": 9, "parts": { "width": 384, "ffn": 1344, "depth": 9, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8・9層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "共通で混合", "reference_layers": [ [ 8, 9 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "中止・削除済み（ユーザー指定）", "architecture": { "id": "VISTARC_CANDIDATE_READER", "width": 384, "layers": 9, "hidden": 1344, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_last2", "attention_received_gate": true, "output_components": 3, "output_state_layers": 2, "reference_attention_width": 256, "ffn_additive_bypass": false, "ffn_input_output_gates": false }, "evaluation": { "official_repo": "babylm-org/babylm-eval", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "track": "strict-small", "backend": "causal", "final": [ "local six axes only" ], "intermediate": "No evaluation until100M training completes", "local_full_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "local_reading_manifest_sha256": "dcaedf9a58ce2fd8f0fb112407dcf9e7fc27d4a156723a27718554384d3e1856", "runtime_policy": "FP32 canonical prefix graph, validated against strict recurrence for this MoS3_LAST2 run; eager fallback on mismatch", "runtime_selection": { "graph_chunk": 4, "max_batch": 1024, "kv_token_budget": 65536, "prefix_graph": true, "fused_mixture": false }, "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "objective": { "primary": "highest_observed_local_six_axis", "additional": "highest_observed_five_axis_without_piqa", "benchmark_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "serial after100M training completes", "no_endpoint_penalty": true, "parent_peak_five_axis": 51.72, "same_family_peak_six_axis": 49.315, "comparison_scope": "last2_256 parent plus interrupted last2_272 and last3/last4 references" }, "milestones_words": [ 60000000, 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "comparison": "all known models and original#1,#2,#3; same ages and independent observed5/6peaks", "supervised7": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "training_precision": "bf16_autocast_fp32_master", "evaluation_precision": "fp32", "tf32": true, "deterministic": true, "vram_budget_gib": 22, "compile_threads": 2, "compile_fullgraph": true, "compile_dynamic": false, "power_limit_watts": 260, "recompile_limit": 32, "fallback_random": true }, "condition_conflicts": [], "profile": 148, "sources": [ "sec08ce221ab9dc" ], "points": [], "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 280, "id": "run/IPRLM384_MODEL277_DEPTH7_FFN1408_SEED20260907_100M_20261028", "original_name": "IPRLM384_MODEL277_DEPTH7_FFN1408_SEED20260907_100M_20261028", "title": "384幅・FFN1408・7層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 7, "parts": { "width": 384, "ffn": 1408, "depth": 7, "attention": "12×32", "router": 32, "candidate": "未確認", "reader": 48, "received": "あり", "output": "MoS3", "reference": "7層 / 6・7層 / 5・6・7層", "ffn_add": "未確認", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 7 ], [ 6, 7 ], [ 5, 6, 7 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 7, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 7 ], [ 6, 7 ], [ 5, 6, 7 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 91, "plastic_rank": 16, "plastic_state": "owner-local F", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9 }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "a8280b292806610a2e7d5e3c9d21dd42583776bda888159938666b5202177d25", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "Allknownmodels; parent277and270, F10L276, plain8L275 and#3depthseries", "supervised7": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source_v2", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "all+outer_cuda_graph; compiled forward/backward only, AdamW/clip and metadata validation outside capture", "cuda_graph_origin": "#277 corrected GraphUpdate, unchanged kernel; fresh7L shape/Adam/RNG/padded-tail timing before training" }, "condition_conflicts": [], "profile": 128, "sources": [ "s98820079b3ebf0" ], "points": [ { "age": 70.0, "six": 47.59, "five": 49.966, "tasks": { "blimp_filtered": 68.56, "supplement_filtered": 56.39, "ewok_filtered": 51.39, "entity_tracking": 20.9, "comps": 52.59, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 52.0 }, "checkpoint": "0dc19ba86202466cefbbe824283201ee2e9c08dad6d3188088b33dfbc08be0d8", "precision": "FP32", "verification": "primary", "sources": [ "s7368e7bb5714f9", "s3aced3f243399b" ] }, { "age": 80.0, "six": 48.02166666666667, "five": 50.584, "tasks": { "blimp_filtered": 67.94, "supplement_filtered": 59.27, "ewok_filtered": 51.68, "entity_tracking": 21.62, "comps": 52.41, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 51.0 }, "checkpoint": "adcab8a45327986894634a0aff416853ac015be6a5420ed6815fe81ef1df353f", "precision": "FP32", "verification": "primary", "sources": [ "s953e054eb6d39c", "s3aced3f243399b" ] }, { "age": 90.0, "six": 47.9875, "five": 49.658, "tasks": { "blimp_filtered": 67.41, "supplement_filtered": 57.54, "ewok_filtered": 50.91, "entity_tracking": 20.11, "comps": 52.32, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 55.0 }, "checkpoint": "00803bed90a2c599cd93f811427cf1a4bd65c65aaecb36c4e003336edd959d0f", "precision": "FP32", "verification": "primary", "sources": [ "s2ba660c3e3d96f", "s3aced3f243399b" ] }, { "age": 100.0, "six": 47.5375, "five": 50.312, "tasks": { "blimp_filtered": 67.14, "supplement_filtered": 58.13, "ewok_filtered": 52.01, "entity_tracking": 22.01, "comps": 52.27, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 45.0 }, "checkpoint": "2dae2f8e1e4d01888d38f998034c685cee3aed4fe898268fafc85804a81f13b2", "precision": "FP32", "verification": "primary", "sources": [ "s9e96bfe7383a1c", "s3aced3f243399b" ] } ], "peak": { "five": { "score": 50.584, "ages": [ 80.0 ], "rank": 102 }, "six": { "score": 48.02166666666667, "ages": [ 80.0 ], "rank": 117 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 281, "id": "run/IPRLM384_MODEL277_FFN_F_LR110_SEED20260907_100M_20261028", "original_name": "IPRLM384_MODEL277_FFN_F_LR110_SEED20260907_100M_20261028", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "未確認", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "未確認", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 104, "plastic_rank": 16, "plastic_state": "owner-local F", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9 }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "a8280b292806610a2e7d5e3c9d21dd42583776bda888159938666b5202177d25", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "parent#277,all existing model inventory; same exact recurrent FP32 evaluation", "supervised7": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "all+outer_cuda_graph; compiled forward/backward only, AdamW/clip and metadata validation outside capture", "cuda_graph_origin": "Identical accepted #277 graph_update.py; changed optimizer groups verified before training" }, "condition_conflicts": [], "profile": 70, "sources": [ "sb4966a088ee5dc" ], "points": [ { "age": 70.0, "six": 48.87833333333333, "five": 50.636, "tasks": { "blimp_filtered": 69.28, "supplement_filtered": 59.43, "ewok_filtered": 50.93, "entity_tracking": 21.08, "comps": 52.46, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 53.0 }, "checkpoint": "cd94108646a5e86148ca4e71b71dd50d4624765af0d838235879d18c3793fef0", "precision": "FP32", "verification": "primary", "sources": [ "sb7f3dc20dfbd59", "s53626f4273513e" ] }, { "age": 80.0, "six": 49.5675, "five": 50.86, "tasks": { "blimp_filtered": 68.98, "supplement_filtered": 60.96, "ewok_filtered": 50.68, "entity_tracking": 20.7, "comps": 52.98, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 60.0 }, "checkpoint": "43ed35f379e18bbd1362fcd7c8723003f2a07ce187be74bc1ac1f198588e0e03", "precision": "FP32", "verification": "primary", "sources": [ "sdb1c82b3fbe742", "s53626f4273513e" ] }, { "age": 90.0, "six": 48.439166666666665, "five": 50.4, "tasks": { "blimp_filtered": 69.21, "supplement_filtered": 59.2, "ewok_filtered": 52.15, "entity_tracking": 18.98, "comps": 52.46, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 53.0 }, "checkpoint": "37e27424571615e17d17c694256640b995f7404e30b0b9fd8583a5fd0931edad", "precision": "FP32", "verification": "primary", "sources": [ "s7e6fb7c2d35f25", "s53626f4273513e" ] }, { "age": 100.0, "six": 48.17666666666667, "five": 50.676, "tasks": { "blimp_filtered": 68.49, "supplement_filtered": 59.99, "ewok_filtered": 51.92, "entity_tracking": 20.25, "comps": 52.73, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 50.0 }, "checkpoint": "df6ba4f522019998acc126ff041a99242ac486e175a97a0b24d2b961c3af8469", "precision": "FP32", "verification": "primary", "sources": [ "sc670ec8ddeb02c", "s53626f4273513e" ] } ], "peak": { "five": { "score": 50.86, "ages": [ 80.0 ], "rank": 67 }, "six": { "score": 49.5675, "ages": [ 80.0 ], "rank": 5 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 282, "id": "run/IPRLM384_MODEL277_ATTN_F_GATE_LR90_SEED20260907_100M_20261028", "original_name": "IPRLM384_MODEL277_ATTN_F_GATE_LR90_SEED20260907_100M_20261028", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "未確認", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "未確認", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 104, "plastic_rank": 16, "plastic_state": "owner-local F", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9 }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "a8280b292806610a2e7d5e3c9d21dd42583776bda888159938666b5202177d25", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "parent#277,#281 and all existing model inventory; same exact recurrent FP32 evaluation", "supervised7": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "all+outer_cuda_graph; compiled forward/backward only, AdamW/clip and metadata validation outside capture", "cuda_graph_origin": "Identical accepted #277 graph_update.py; changed optimizer groups verified before training" }, "condition_conflicts": [], "profile": 127, "sources": [ "sbca4da26af507b" ], "points": [ { "age": 70.0, "six": 48.56916666666667, "five": 50.85, "tasks": { "blimp_filtered": 69.37, "supplement_filtered": 60.64, "ewok_filtered": 52.03, "entity_tracking": 20.05, "comps": 52.16, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "checkpoint": "020f4488d74f664deddafe585413fefcaf10c00d2673aeb1646fac2ebcef00fa", "precision": "FP32", "verification": "primary", "sources": [ "s0bd2f5d4e48c0c", "s97b06b127a0c78" ] }, { "age": 80.0, "six": 48.60666666666667, "five": 50.992, "tasks": { "blimp_filtered": 68.18, "supplement_filtered": 62.61, "ewok_filtered": 51.41, "entity_tracking": 20.44, "comps": 52.32, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 52.0 }, "checkpoint": "4461e22f3cbc4e6c882fc859a8f63b3c9d33c001461a9303abc4f123133db51c", "precision": "FP32", "verification": "primary", "sources": [ "sca4d28c3a0ccd7", "s97b06b127a0c78" ] }, { "age": 90.0, "six": 48.553333333333335, "five": 50.528, "tasks": { "blimp_filtered": 69.27, "supplement_filtered": 58.61, "ewok_filtered": 51.1, "entity_tracking": 21.35, "comps": 52.31, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 56.0 }, "checkpoint": "27ea6098b0203db611abbf574ab169f48e71b4a18e11c0d9b1d70d41c84c916d", "precision": "FP32", "verification": "primary", "sources": [ "sbd07eb5b2aab9f", "s97b06b127a0c78" ] }, { "age": 100.0, "six": 48.07833333333333, "five": 50.57, "tasks": { "blimp_filtered": 67.97, "supplement_filtered": 60.78, "ewok_filtered": 51.67, "entity_tracking": 20.57, "comps": 51.86, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 46.0 }, "checkpoint": "5fa1fe16c6fcc8e5fcbad19c11c487de3ad06e74ba8f48b6e426695d1f58a292", "precision": "FP32", "verification": "primary", "sources": [ "s6d1e77ce33bb53", "s97b06b127a0c78" ] } ], "peak": { "five": { "score": 50.992, "ages": [ 80.0 ], "rank": 47 }, "six": { "score": 48.60666666666667, "ages": [ 80.0 ], "rank": 69 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 283, "id": "run/IPRLM384_MODEL277_SHARED384_F_CURRENT_ONLY_SEED20260907_100M_20261028", "original_name": "IPRLM384_MODEL277_SHARED384_F_CURRENT_ONLY_SEED20260907_100M_20261028", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "未確認", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "未確認", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL277_SHARED384_CURRENT_ONLY_F", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "initialization_depth_anchor": 9, "F_mode": "direct384", "F_history": false, "F_key_width": 0, "F_outputs": [ "Q", "received_gate", "FFN_input" ], "parameter_count": 62388117 }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "f974cc67714a0096abe19f992dd86c9dc760723d74a18f78b9dc2fb2fccb8338", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "parent#277,#281,#282 and complete model inventory", "supervised7": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "fullcompile+outer CUDAgraph; child numerical/VRAM/speed gate required", "cuda_graph_origin": "Parent277 graph implementation; child full384 equations separately verified" }, "condition_conflicts": [], "profile": 70, "sources": [ "s239dd4cff46b9d" ], "points": [ { "age": 70.0, "six": 47.83583333333333, "five": 50.47, "tasks": { "blimp_filtered": 69.94, "supplement_filtered": 58.75, "ewok_filtered": 51.52, "entity_tracking": 19.3, "comps": 52.84, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 47.0 }, "checkpoint": "5342aee4a93c7b175f5583d3bc54228f78b33015808c47a32d5a02ee05cb6fe1", "precision": "FP32", "verification": "primary", "sources": [ "saba0ca76863c53", "s79ce3e1b972aa2" ] }, { "age": 80.0, "six": 48.348333333333336, "five": 50.288, "tasks": { "blimp_filtered": 67.96, "supplement_filtered": 59.56, "ewok_filtered": 51.85, "entity_tracking": 19.52, "comps": 52.55, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 54.0 }, "checkpoint": "c972eb683bfab0e7ec4f6a45eae2409553e41a85016425a620b36de42f7615dd", "precision": "FP32", "verification": "primary", "sources": [ "s63544f11ecb192", "s79ce3e1b972aa2" ] }, { "age": 90.0, "six": 47.275, "five": 49.894, "tasks": { "blimp_filtered": 67.84, "supplement_filtered": 58.09, "ewok_filtered": 52.15, "entity_tracking": 19.09, "comps": 52.3, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 47.0 }, "checkpoint": "c0aa9c52b36614a97f60b4fe9bae49dfadc0b42134f429bb5fd0407117728617", "precision": "FP32", "verification": "primary", "sources": [ "sbd2ef1605858ba", "s79ce3e1b972aa2" ] }, { "age": 100.0, "six": 47.619166666666665, "five": 49.716, "tasks": { "blimp_filtered": 67.94, "supplement_filtered": 58.41, "ewok_filtered": 50.91, "entity_tracking": 18.85, "comps": 52.47, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 50.0 }, "checkpoint": "bbfae8f4f0860586ba084cd867f968a6e83af72f43fdff2b7c10686bf6ee3e07", "precision": "FP32", "verification": "primary", "sources": [ "sc7a4b5a5cc2ec4", "s79ce3e1b972aa2" ] } ], "peak": { "five": { "score": 50.47, "ages": [ 70.0 ], "rank": 113 }, "six": { "score": 48.348333333333336, "ages": [ 80.0 ], "rank": 91 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 284, "original_name": "IPRLM384_MODEL283_DIRECT_F_WD095_SEED20260907_100M_20261028", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "seed": 20260907, "cohort": "baby", "state": "ユーザー指定で停止・削除", "id": "run/IPRLM384_MODEL283_DIRECT_F_WD095_SEED20260907_100M_20261028", "present": false, "points": [], "parts": {}, "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "profile": null, "sources": [] }, { "number": 285, "original_name": "IPRLM384_MODEL283_DIRECT_F_WD105_SEED20260907_100M_20261028", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "seed": 20260907, "cohort": "baby", "state": "ユーザー指定で停止・削除", "id": "run/IPRLM384_MODEL283_DIRECT_F_WD105_SEED20260907_100M_20261028", "present": false, "points": [], "parts": {}, "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "profile": null, "sources": [] }, { "number": 286, "original_name": "IPRLM384_MODEL283_DIRECT_F_DELTA_INIT105_SEED20260907_100M_20261028", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "seed": 20260907, "cohort": "baby", "state": "ユーザー指定で停止・削除", "id": "run/IPRLM384_MODEL283_DIRECT_F_DELTA_INIT105_SEED20260907_100M_20261028", "present": false, "points": [], "parts": {}, "peak": { "five": { "score": null, "ages": [] }, "six": { "score": null, "ages": [] } }, "clear_points": [], "clear_peak": null, "profile": null, "sources": [] }, { "number": 287, "id": "run/IPRLM384_MODEL283_FFN_CONTEXT384_SEED20260907_20261029", "original_name": "IPRLM384_MODEL283_FFN_CONTEXT384_SEED20260907_20261029", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "未確認", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "未確認", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL277_SHARED384_CURRENT_ONLY_F", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "initialization_depth_anchor": 9, "F_mode": "direct384", "F_history": false, "F_key_width": 0, "F_outputs": [ "Q", "received_gate", "FFN_input" ], "parameter_count": 63567765 }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "f974cc67714a0096abe19f992dd86c9dc760723d74a18f78b9dc2fb2fccb8338", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "parent#283 and historical#277;full model inventory", "supervised7": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "fullcompile+outerCUDAgraph;child GPU acceptance required", "cuda_graph_origin": "Parent283 existing implementation;newcontext nonzero-tested before launch" }, "condition_conflicts": [], "profile": 70, "sources": [ "s3c4f540e66e34f" ], "points": [ { "age": 70.0, "six": 47.93333333333333, "five": 50.172, "tasks": { "blimp_filtered": 68.66, "supplement_filtered": 58.34, "ewok_filtered": 50.86, "entity_tracking": 20.33, "comps": 52.67, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 56.0 }, "checkpoint": "356b42326da4d0a2dfcd8206398628b10123bcf6aaf61c39a721baf8baf17646", "precision": "FP32", "verification": "primary", "sources": [ "s6183aa6dacb1d2", "s3aee8871fec2e0" ] }, { "age": 80.0, "six": 48.343333333333334, "five": 50.476, "tasks": { "blimp_filtered": 68.69, "supplement_filtered": 60.81, "ewok_filtered": 49.84, "entity_tracking": 20.59, "comps": 52.45, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 54.0 }, "checkpoint": "ff26d137a6a8238b31c3ed587cd5052de6659772fe51c7d393b5a3fe9310e269", "precision": "FP32", "verification": "primary", "sources": [ "s7038f6d04a3c48", "s3aee8871fec2e0" ] }, { "age": 90.0, "six": 47.634166666666665, "five": 49.528, "tasks": { "blimp_filtered": 68.31, "supplement_filtered": 56.35, "ewok_filtered": 50.57, "entity_tracking": 20.21, "comps": 52.2, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 54.0 }, "checkpoint": "15c1f0f1a2b547867595ef2e44db89a697b3a266f48a73bfca0580a0b46c6692", "precision": "FP32", "verification": "primary", "sources": [ "s115e35affb82c2", "s3aee8871fec2e0" ] }, { "age": 100.0, "six": 47.11833333333333, "five": 49.806, "tasks": { "blimp_filtered": 68.29, "supplement_filtered": 58.0, "ewok_filtered": 50.03, "entity_tracking": 20.35, "comps": 52.36, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 46.0 }, "checkpoint": "47436472861f557c3ddd39f2cc79d34aae76a3dd10bb3272f008f55a59c8c87b", "precision": "FP32", "verification": "primary", "sources": [ "s9489619faa629c", "s3aee8871fec2e0" ] } ], "peak": { "five": { "score": 50.476, "ages": [ 80.0 ], "rank": 110 }, "six": { "score": 48.343333333333334, "ages": [ 80.0 ], "rank": 93 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 288, "id": "run/IPRLM384_MODEL277_F_RANK32_SEED20260907_20261029", "original_name": "IPRLM384_MODEL277_F_RANK32_SEED20260907_20261029", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "未確認", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "未確認", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 104, "plastic_rank": 32, "plastic_state": "owner-local F", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9 }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "a8280b292806610a2e7d5e3c9d21dd42583776bda888159938666b5202177d25", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "#277 rank16 parent and full fixed-number inventory", "supervised7": false, "authorization": "User GO on2026-10-29 after implementation of#277 oldF rank32; same parent100M recipe and70/80/90/100M local6" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_PROBE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "all+outer_cuda_graph; compiled forward/backward only, AdamW/clip and metadata validation outside capture", "cuda_graph_origin": "#277 source/graph_update.py; rank32 GPU validation pending" }, "condition_conflicts": [], "profile": 70, "sources": [ "s24cd89f338e52a" ], "points": [ { "age": 70.0, "six": 48.465, "five": 50.342, "tasks": { "blimp_filtered": 69.87, "supplement_filtered": 58.57, "ewok_filtered": 51.1, "entity_tracking": 20.0, "comps": 52.17, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 50.0 }, "checkpoint": "d781cda48fc6c2286921fda137c152833e65a660cac405c9dc5faed84395dcbe", "precision": "FP32", "verification": "primary", "sources": [ "s94ddf0a21e7b23", "s41cb5d040d710a" ] }, { "age": 80.0, "six": 48.336666666666666, "five": 50.674, "tasks": { "blimp_filtered": 68.5, "supplement_filtered": 60.99, "ewok_filtered": 51.73, "entity_tracking": 19.72, "comps": 52.43, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "f357436da7e796598f1b24e394935e4138b6fe3612b86f2e4660ab8fe2480e07", "precision": "FP32", "verification": "primary", "sources": [ "s99bceb12b5a22a", "s41cb5d040d710a" ] }, { "age": 90.0, "six": 47.54416666666667, "five": 50.208, "tasks": { "blimp_filtered": 68.0, "supplement_filtered": 59.89, "ewok_filtered": 51.9, "entity_tracking": 19.11, "comps": 52.14, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 50.0 }, "checkpoint": "56a8e9c752c4e252c0f76fdc1e8b40a46193527dd8047f5bca44b5a04568ead4", "precision": "FP32", "verification": "primary", "sources": [ "s1da953a4f400fb", "s41cb5d040d710a" ] }, { "age": 100.0, "six": 48.358333333333334, "five": 50.806, "tasks": { "blimp_filtered": 69.12, "supplement_filtered": 60.1, "ewok_filtered": 52.39, "entity_tracking": 19.88, "comps": 52.54, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 47.0 }, "checkpoint": "52c69fb8e654c02932fb7e4b64ee0f98b99a880431afdf5a012bb213336c8e48", "precision": "FP32", "verification": "primary", "sources": [ "s1169ffcfbbcc51", "s41cb5d040d710a" ] } ], "peak": { "five": { "score": 50.806, "ages": [ 100.0 ], "rank": 76 }, "six": { "score": 48.465, "ages": [ 70.0 ], "rank": 81 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 289, "id": "run/IPRLM384_MODEL277_F_VALUE32_KEY16_SEED20260907_20261029", "original_name": "IPRLM384_MODEL277_F_VALUE32_KEY16_SEED20260907_20261029", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "未確認", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "未確認", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 104, "plastic_rank": 32, "plastic_state": "owner-local F", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9, "plastic_value_rank": 32, "plastic_key_rank": 16, "plastic_memory_shape": [ 32, 16 ] }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "a8280b292806610a2e7d5e3c9d21dd42583776bda888159938666b5202177d25", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "#277 rank16 parent and full fixed-number inventory", "supervised7": false, "authorization": "2026-10-29 user: first error/state32 observer16, next error16 observer32, finally32/32; continue prior GO scope0M100M and70/80/90/100M local6, same parent277 LR and training." }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_PROBE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "all+outer_cuda_graph; compiled forward/backward only, AdamW/clip and metadata validation outside capture", "cuda_graph_origin": "#277 source/graph_update.py; rectangular GPU validation required" }, "condition_conflicts": [], "profile": 70, "sources": [ "s007a128c5c5463" ], "points": [ { "age": 70.0, "six": 48.3275, "five": 50.266, "tasks": { "blimp_filtered": 69.12, "supplement_filtered": 59.63, "ewok_filtered": 51.1, "entity_tracking": 19.47, "comps": 52.01, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 53.0 }, "checkpoint": "887f5d5b074e033710b919490f73a4f078e1ee01ed041172e1ac0de0f73d5566", "precision": "FP32", "verification": "primary", "sources": [ "s626d896a208a9d", "s2abcdfee8af6f6" ] }, { "age": 80.0, "six": 48.568333333333335, "five": 50.358, "tasks": { "blimp_filtered": 68.57, "supplement_filtered": 60.2, "ewok_filtered": 51.56, "entity_tracking": 19.02, "comps": 52.44, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 54.0 }, "checkpoint": "956a294615174f2e677e485c218739da73add5db2ccbe382e7bf4d8f8472e727", "precision": "FP32", "verification": "primary", "sources": [ "s4b5589228cf2de", "s2abcdfee8af6f6" ] }, { "age": 90.0, "six": 46.85916666666667, "five": 49.492, "tasks": { "blimp_filtered": 67.75, "supplement_filtered": 57.7, "ewok_filtered": 50.22, "entity_tracking": 20.02, "comps": 51.77, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 47.0 }, "checkpoint": "c448997ed2cc9465af6a4a599033d05b75ddeb9cb33bbb0a55a642d14a1683d4", "precision": "FP32", "verification": "primary", "sources": [ "s00bb9f6261e829", "s2abcdfee8af6f6" ] }, { "age": 100.0, "six": 47.75833333333333, "five": 49.88, "tasks": { "blimp_filtered": 68.72, "supplement_filtered": 57.42, "ewok_filtered": 50.73, "entity_tracking": 20.51, "comps": 52.02, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 51.0 }, "checkpoint": "7a243ef6670f4b148abae5bd2d8208a6c481bb31fcf78dddcb9e414187c482be", "precision": "FP32", "verification": "primary", "sources": [ "sde7b1be04b3857", "s2abcdfee8af6f6" ] } ], "peak": { "five": { "score": 50.358, "ages": [ 80.0 ], "rank": 122 }, "six": { "score": 48.568333333333335, "ages": [ 80.0 ], "rank": 72 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 290, "id": "run/IPRLM384_MODEL277_F_VALUE16_KEY32_SEED20260907_20261029", "original_name": "IPRLM384_MODEL277_F_VALUE16_KEY32_SEED20260907_20261029", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "未確認", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "未確認", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 104, "plastic_rank": 16, "plastic_state": "owner-local F", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9, "plastic_value_rank": 16, "plastic_key_rank": 32, "plastic_memory_shape": [ 16, 32 ] }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "a8280b292806610a2e7d5e3c9d21dd42583776bda888159938666b5202177d25", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "#277 rank16 parent and full fixed-number inventory", "supervised7": false, "authorization": "2026-10-29 user: first error/state32 observer16, next error16 observer32, finally32/32; continue prior GO scope0M100M and70/80/90/100M local6, same parent277 LR and training." }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_PROBE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "all+outer_cuda_graph; compiled forward/backward only, AdamW/clip and metadata validation outside capture", "cuda_graph_origin": "#277 source/graph_update.py; rectangular GPU validation required" }, "condition_conflicts": [], "profile": 70, "sources": [ "s243aa778b9c1bb" ], "points": [ { "age": 70.0, "six": 48.836666666666666, "five": 50.868, "tasks": { "blimp_filtered": 69.91, "supplement_filtered": 59.34, "ewok_filtered": 50.9, "entity_tracking": 21.3, "comps": 52.89, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 56.0 }, "checkpoint": "8578862d7e29f30cb7c2dc2b3723f1a76560301513a68e1060aea366571f6f1c", "precision": "FP32", "verification": "primary", "sources": [ "sdffcb040db98cb", "s31f7596ec7f603" ] }, { "age": 80.0, "six": 49.23166666666667, "five": 51.642, "tasks": { "blimp_filtered": 69.55, "supplement_filtered": 62.1, "ewok_filtered": 52.46, "entity_tracking": 21.55, "comps": 52.55, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 53.0 }, "checkpoint": "ee9b875d48e25e94e3a7c914fd4226fa22abbc2cb9e34d7fe6e38bb9e78a40d7", "precision": "FP32", "verification": "primary", "sources": [ "s1b0cc531280bb0", "s31f7596ec7f603" ] }, { "age": 90.0, "six": 47.916666666666664, "five": 50.758, "tasks": { "blimp_filtered": 69.23, "supplement_filtered": 58.13, "ewok_filtered": 52.94, "entity_tracking": 21.1, "comps": 52.39, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 48.0 }, "checkpoint": "e5b4ea4f72fc6430725d759b37b520f5df8153035d887d3885c8d61aa8191d77", "precision": "FP32", "verification": "primary", "sources": [ "s0de4e4d4cc9a06", "s31f7596ec7f603" ] }, { "age": 100.0, "six": 48.525, "five": 50.706, "tasks": { "blimp_filtered": 68.89, "supplement_filtered": 60.51, "ewok_filtered": 51.84, "entity_tracking": 20.1, "comps": 52.19, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 50.0 }, "checkpoint": "9497b56ef31d5363314daa2c0dd5d82d0b29ebbed950cc65318183006c7d03a0", "precision": "FP32", "verification": "primary", "sources": [ "s91c64d2e3af2c9", "s31f7596ec7f603" ] } ], "peak": { "five": { "score": 51.642, "ages": [ 80.0 ], "rank": 6 }, "six": { "score": 49.23166666666667, "ages": [ 80.0 ], "rank": 17 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 291, "id": "run/IPRLM384_MODEL277_F16_NO_HISTORY_SEED20260907_20261029", "original_name": "IPRLM384_MODEL277_F16_NO_HISTORY_SEED20260907_20261029", "title": "384幅・FFN1408・8層・F16×16履歴なし", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "候補平均・分散を帰還／Fは全候補平均・現在分のみ", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "F_value_width": 16, "F_key_width": 16, "F_history": false, "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 104, "plastic_rank": 16, "plastic_state": "current outer product only; not accumulated/stored", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9, "F_variant": "no_history", "F_history": false, "F_key_width": 16, "F_value_width": 16, "F_threshold": false, "plastic_value_rank": 16, "plastic_key_rank": 16, "plastic_memory_shape": [ 16, 16 ], "ffn_additive_bypass": false, "ffn_gates": false }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "e78cb5b9ee53d7e749d600f9ad53f0809cb6528a1963e1af5445ad3d775732bb", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "After100M training, score each saved checkpoint", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "same-seed #277 plus all-model inventory when scored", "supervised7": false, "authorization": "User2026-10-29: #291 and #292 automatic sequential GO. Same preceding #277-family100M recipe and70/80/90/100M local6 retained; no extra seed or finetune7." }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json", "manifest_scope": "Shared exact parent files; small manifest/tokenizer hashes checked; no fresh repacking or full corpus rescan" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "all+outer_cuda_graph; compiled forward/backward only, AdamW/clip and metadata validation outside capture", "cuda_graph_origin": "#277 existing full compile+CUDA Graph, short native/update/rebuild/speed tests required", "acceptance": "New child GPU tests required before production" }, "condition_conflicts": [], "profile": 70, "sources": [ "s18a5e6a870fb7e" ], "points": [ { "age": 70.0, "six": 48.5, "five": 50.576, "tasks": { "blimp_filtered": 69.91, "supplement_filtered": 58.32, "ewok_filtered": 52.2, "entity_tracking": 20.24, "comps": 52.21, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 51.0 }, "checkpoint": "ed1ac3933ac35e53d827ca954dafc2f804acfff2a91759054ff264d64986b0d4", "precision": "FP32", "verification": "primary", "sources": [ "s1a52a8ba6063eb", "sd74cb039b4455c" ] }, { "age": 80.0, "six": 48.42583333333334, "five": 50.684, "tasks": { "blimp_filtered": 69.06, "supplement_filtered": 60.69, "ewok_filtered": 51.71, "entity_tracking": 19.78, "comps": 52.18, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 50.0 }, "checkpoint": "6653f16c5b1270d4aae56f05081e20df1993af49468e40b1a0224897c366ee30", "precision": "FP32", "verification": "primary", "sources": [ "sbd26bc9f053cb1", "sd74cb039b4455c" ] }, { "age": 90.0, "six": 48.083333333333336, "five": 50.57, "tasks": { "blimp_filtered": 68.37, "supplement_filtered": 60.61, "ewok_filtered": 51.26, "entity_tracking": 20.89, "comps": 51.72, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 48.0 }, "checkpoint": "9e00af4273827c32bc683c766191c40232dbc03591226a83c2e868c1ff086a2c", "precision": "FP32", "verification": "primary", "sources": [ "s175a9088739f45", "sd74cb039b4455c" ] }, { "age": 100.0, "six": 48.12166666666667, "five": 50.122, "tasks": { "blimp_filtered": 68.22, "supplement_filtered": 59.31, "ewok_filtered": 51.5, "entity_tracking": 20.05, "comps": 51.53, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 51.0 }, "checkpoint": "99419ecc678c9b5d14a13d0f986d5c66d41f93eaed181d2a1f22c3fc543bd326", "precision": "FP32", "verification": "primary", "sources": [ "sa0030993030339", "sd74cb039b4455c" ] } ], "peak": { "five": { "score": 50.684, "ages": [ 80.0 ], "rank": 84 }, "six": { "score": 48.5, "ages": [ 70.0 ], "rank": 77 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 292, "id": "run/IPRLM384_MODEL277_F16_THRESHOLD_SEED20260907_20261029", "original_name": "IPRLM384_MODEL277_F16_THRESHOLD_SEED20260907_20261029", "title": "384幅・FFN1408・8層・F16×16しきい値候補", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "候補平均・分散を帰還／Fはしきい値候補平均・履歴あり", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "F_value_width": 16, "F_key_width": 16, "F_history": true, "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 104, "plastic_rank": 16, "plastic_state": "owner-local recurrent F", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9, "F_variant": "threshold", "F_history": true, "F_key_width": 16, "F_value_width": 16, "F_threshold": true, "plastic_value_rank": 16, "plastic_key_rank": 16, "plastic_memory_shape": [ 16, 16 ], "ffn_additive_bypass": false, "ffn_gates": false }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "e78cb5b9ee53d7e749d600f9ad53f0809cb6528a1963e1af5445ad3d775732bb", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "After100M training, score each saved checkpoint", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "same-seed #277 plus all-model inventory when scored", "supervised7": false, "authorization": "User2026-10-29: #291 and #292 automatic sequential GO. Same preceding #277-family100M recipe and70/80/90/100M local6 retained; no extra seed or finetune7." }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json", "manifest_scope": "Shared exact parent files; small manifest/tokenizer hashes checked; no fresh repacking or full corpus rescan" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "all+outer_cuda_graph; compiled forward/backward only, AdamW/clip and metadata validation outside capture", "cuda_graph_origin": "#277 existing full compile+CUDA Graph, short native/update/rebuild/speed tests required", "acceptance": "New child GPU tests required before production" }, "condition_conflicts": [], "profile": 70, "sources": [ "se9051413196e86" ], "points": [ { "age": 70.0, "six": 47.53666666666667, "five": 49.908, "tasks": { "blimp_filtered": 68.56, "supplement_filtered": 58.35, "ewok_filtered": 52.1, "entity_tracking": 18.77, "comps": 51.76, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 50.0 }, "checkpoint": "df2c1ef01c6315e816c92fe872447dc007ab527765678b47d4a2ac5f2cefc45d", "precision": "FP32", "verification": "primary", "sources": [ "s1e1010cd82d246", "s6322c13566879d" ] }, { "age": 80.0, "six": 48.73166666666667, "five": 50.554, "tasks": { "blimp_filtered": 67.85, "supplement_filtered": 60.62, "ewok_filtered": 52.68, "entity_tracking": 19.59, "comps": 52.03, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 54.0 }, "checkpoint": "76171621cde3596bb96b1957234a8743ca9bd0e9d515d6fe04bcb0f16f3e2890", "precision": "FP32", "verification": "primary", "sources": [ "s1bb4cbe7f0ffb6", "s6322c13566879d" ] }, { "age": 90.0, "six": 47.66166666666667, "five": 49.958, "tasks": { "blimp_filtered": 67.57, "supplement_filtered": 59.25, "ewok_filtered": 51.52, "entity_tracking": 19.33, "comps": 52.12, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 51.0 }, "checkpoint": "b7eb36113ca769ddb1136fae419fdc3451060764e7ad7ac77d454affc5b51069", "precision": "FP32", "verification": "primary", "sources": [ "s2d0128688dd713", "s6322c13566879d" ] }, { "age": 100.0, "six": 48.1875, "five": 50.592, "tasks": { "blimp_filtered": 68.93, "supplement_filtered": 59.96, "ewok_filtered": 52.32, "entity_tracking": 19.7, "comps": 52.05, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 50.0 }, "checkpoint": "6cf25050cdc318cc42af243a4efbc42bf47087d21ad4942d9b3fc42c9ffadd65", "precision": "FP32", "verification": "primary", "sources": [ "sdd8a576c5cbd18", "s6322c13566879d" ] } ], "peak": { "five": { "score": 50.592, "ages": [ 100.0 ], "rank": 98 }, "six": { "score": 48.73166666666667, "ages": [ 80.0 ], "rank": 58 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 293, "id": "run/IPRLM384_MODEL277_F16_GATED_DELTA_SEED20260907_20261029", "original_name": "IPRLM384_MODEL277_F16_GATED_DELTA_SEED20260907_20261029", "title": "384幅・FFN1408・8層・F16×16 GDN差分更新・履歴あり", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "候補平均・分散を帰還／Fは全候補平均・GDN差分更新・履歴あり", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "F_history": true, "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 104, "plastic_rank": 16, "plastic_state": "16x16 gated-delta associative prediction-error history", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9, "F_variant": "gated_delta", "F_history": true, "F_key_width": 16, "F_value_width": 16, "F_threshold": false, "plastic_value_rank": 16, "plastic_key_rank": 16, "plastic_memory_shape": [ 16, 16 ], "ffn_additive_bypass": false, "ffn_gates": false, "F_update": "M=a*F; F=M+b*(v-M@previous_key) outer previous_key" }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "e78cb5b9ee53d7e749d600f9ad53f0809cb6528a1963e1af5445ad3d775732bb", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "After100M training", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "same-seed #277 plus all-model inventory when scored", "supervised7": false, "authorization": "User2026-10-29: implement GDN history with variable writing and automatic GO after291/292, speed-optimize first; same100M and70/80/90/100M recipe" }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json", "manifest_scope": "Shared exact parent files; small manifest/tokenizer hashes checked; no fresh repacking or full corpus rescan" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "Gated delta each arrival", "compiled_source": "source", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "all+outer_cuda_graph; compiled forward/backward only, AdamW/clip and metadata validation outside capture", "cuda_graph_origin": "#277 compiled native+outer CUDA Graph; new exact GDN parallel scan", "acceptance": "GDN GPU parity and measured solver/speed required" }, "condition_conflicts": [], "profile": 129, "sources": [ "sff3f10b0921fdd" ], "points": [ { "age": 70.0, "six": 49.0375, "five": 51.124, "tasks": { "blimp_filtered": 70.35, "supplement_filtered": 59.15, "ewok_filtered": 52.78, "entity_tracking": 21.64, "comps": 51.7, "global_piqa_parallel": 26.21, "global_piqa_nonparallel": 51.0 }, "checkpoint": "e09655eaeaafa155ef0b752660fa08cb5cb678d714228c833847123c5b446f29", "precision": "FP32", "verification": "primary", "sources": [ "scc90771892d5ee", "sf2715dfecc53d1" ] }, { "age": 80.0, "six": 48.60666666666667, "five": 50.998, "tasks": { "blimp_filtered": 68.5, "supplement_filtered": 61.49, "ewok_filtered": 52.81, "entity_tracking": 19.9, "comps": 52.29, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 50.0 }, "checkpoint": "8129610a6f3d1dc665c480e261aba90229f7eed5b44011bd8187088bd59f7f5f", "precision": "FP32", "verification": "primary", "sources": [ "s8b410e94934018", "sf2715dfecc53d1" ] }, { "age": 90.0, "six": 47.979166666666664, "five": 50.142, "tasks": { "blimp_filtered": 69.4, "supplement_filtered": 58.51, "ewok_filtered": 51.32, "entity_tracking": 19.67, "comps": 51.81, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 52.0 }, "checkpoint": "961552d587f7fbc4030106ca80ecdb37cf9cce6e4e34e5d47a2ece3512c95263", "precision": "FP32", "verification": "primary", "sources": [ "sffc2bf409425ea", "sf2715dfecc53d1" ] }, { "age": 100.0, "six": 48.835, "five": 50.578, "tasks": { "blimp_filtered": 69.06, "supplement_filtered": 60.34, "ewok_filtered": 52.34, "entity_tracking": 18.81, "comps": 52.34, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 55.0 }, "checkpoint": "ce48457c375e7c96ceeef2e8ef4d10059c3abdb33c3d3a89ac8e182329688388", "precision": "FP32", "verification": "primary", "sources": [ "s88a1c7773039ed", "sf2715dfecc53d1" ] } ], "peak": { "five": { "score": 51.124, "ages": [ 70.0 ], "rank": 32 }, "six": { "score": 49.0375, "ages": [ 70.0 ], "rank": 28 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 294, "id": "run/IPRLM384_MODEL277_SEED731946281_100M_20261030", "original_name": "IPRLM384_MODEL277_SEED731946281_100M_20261030", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 731946281, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 731946281, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 104, "plastic_rank": 16, "plastic_state": "owner-local F", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9, "ffn_additive_bypass": false, "candidate_handling": "候補の平均・分散を帰還" }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "a8280b292806610a2e7d5e3c9d21dd42583776bda888159938666b5202177d25", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "#277 same-age and observed peaks;#1 and full compatible inventory; keep seed runs distinct", "supervised7": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "all+outer_cuda_graph; compiled forward/backward only, AdamW/clip and metadata validation outside capture", "cuda_graph_origin": "#276/extensions/steady_speed_20261028/graph_update.py; revalidated at8layers before fresh training" }, "condition_conflicts": [], "profile": 70, "sources": [ "s6ca405fde814e0" ], "points": [ { "age": 70.0, "six": 48.4975, "five": 50.47, "tasks": { "blimp_filtered": 69.67, "supplement_filtered": 58.63, "ewok_filtered": 51.72, "entity_tracking": 20.43, "comps": 51.9, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 53.0 }, "checkpoint": "ba1b07fc6f9a03cc0c12356fe1b1fe8a69f845c1104cbaa168cab1f5ffc92c1a", "precision": "FP32", "verification": "primary", "sources": [ "sf4b6831d44d4a6", "s25fe9f3d442b87" ] }, { "age": 80.0, "six": 48.87083333333333, "five": 50.818, "tasks": { "blimp_filtered": 69.04, "supplement_filtered": 60.56, "ewok_filtered": 51.58, "entity_tracking": 20.62, "comps": 52.29, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 54.0 }, "checkpoint": "86951f263853441e16a98f6c091b2fdf24c8426860084c5afca0c2227a0618e6", "precision": "FP32", "verification": "primary", "sources": [ "s676426e28b1282", "s25fe9f3d442b87" ] }, { "age": 90.0, "six": 47.861666666666665, "five": 50.516, "tasks": { "blimp_filtered": 67.82, "supplement_filtered": 58.29, "ewok_filtered": 52.03, "entity_tracking": 21.93, "comps": 52.51, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 42.0 }, "checkpoint": "b569943f0d87915e011821dd45e769728abbee9785d3ef4f178ef2d8f93cc5f1", "precision": "FP32", "verification": "primary", "sources": [ "sa0a61890479318", "s25fe9f3d442b87" ] }, { "age": 100.0, "six": 48.25333333333333, "five": 50.374, "tasks": { "blimp_filtered": 69.56, "supplement_filtered": 57.47, "ewok_filtered": 51.55, "entity_tracking": 20.75, "comps": 52.54, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 52.0 }, "checkpoint": "4e131b380b4ff8dd9d325fc8aa98564879487a29f8b072e096e88bb4612ee2e9", "precision": "FP32", "verification": "primary", "sources": [ "scf5688f2cd97f8", "s25fe9f3d442b87" ] } ], "peak": { "five": { "score": 50.818, "ages": [ 80.0 ], "rank": 74 }, "six": { "score": 48.87083333333333, "ages": [ 80.0 ], "rank": 46 } }, "clear_points": [ { "number": 294, "age": 80, "seed": 731946281, "checkpoint_sha256": "86951f263853441e16a98f6c091b2fdf24c8426860084c5afca0c2227a0618e6", "old_six": 48.87083333333333, "old_five": 50.818, "old_tasks": { "blimp_filtered": 69.04, "supplement_filtered": 60.56, "ewok_filtered": 51.58, "entity_tracking": 20.62, "comps": 52.29, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 54.0 }, "correct": 12171, "total": 17274, "accuracy": 70.45849253212921, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6894, "total": 8392, "ties": 0 }, "comps": { "correct": 3859, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 16, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 0, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1002, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 0, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 3, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 72, "total": 114, "ties": 0 }, "boolq": { "correct": 9, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 232, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 243.60875193899847, "checkpoint_rank": 18, "rank": 15 } ], "clear_peak": { "number": 294, "age": 80, "seed": 731946281, "checkpoint_sha256": "86951f263853441e16a98f6c091b2fdf24c8426860084c5afca0c2227a0618e6", "old_six": 48.87083333333333, "old_five": 50.818, "old_tasks": { "blimp_filtered": 69.04, "supplement_filtered": 60.56, "ewok_filtered": 51.58, "entity_tracking": 20.62, "comps": 52.29, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 54.0 }, "correct": 12171, "total": 17274, "accuracy": 70.45849253212921, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6894, "total": 8392, "ties": 0 }, "comps": { "correct": 3859, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 16, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 0, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 1002, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 0, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 3, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 72, "total": 114, "ties": 0 }, "boolq": { "correct": 9, "total": 53, "ties": 0 }, "mnli": { "correct": 24, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 232, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 243.60875193899847, "checkpoint_rank": 18, "rank": 15 }, "present": true }, { "number": 295, "id": "run/IPRLM256_MODEL277_FFN896_SEED20260907_100M_20261030", "original_name": "IPRLM256_MODEL277_FFN896_SEED20260907_100M_20261030", "title": "256幅・FFN896・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 256, "ffn": 896, "depth": 8, "parts": { "width": 256, "ffn": 896, "depth": 8, "attention": "8×32", "router": 32, "candidate": "未確認", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "未確認", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.5 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 256, "layers": 8, "hidden": 896, "attention_heads": 8, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 20260907, "attention_width": 256, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 72, "plastic_rank": 16, "plastic_state": "owner-local F", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9 }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "a8280b292806610a2e7d5e3c9d21dd42583776bda888159938666b5202177d25", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 70000000, 80000000, 90000000, 100000000 ], "timing": "after100M", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "#277/#294, old256 #11/#13/#20/#57, #1/#81,all compatible inventory; observed peaks primary,matched ages secondary", "supervised7": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "all+outer_cuda_graph; compiled forward/backward only, AdamW/clip and metadata validation outside capture", "cuda_graph_origin": "#276/extensions/steady_speed_20261028/graph_update.py; revalidated at8layers before fresh training" }, "condition_conflicts": [], "profile": 70, "sources": [ "sb8782c96106703" ], "points": [ { "age": 70.0, "six": 48.505, "five": 51.064, "tasks": { "blimp_filtered": 70.09, "supplement_filtered": 58.36, "ewok_filtered": 52.32, "entity_tracking": 22.35, "comps": 52.2, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 52.0 }, "checkpoint": "0969c542f9f0f4056718e3973b77afe56160f1cf6fe5bc574ba83d37b1bc5786", "precision": "FP32", "verification": "primary", "sources": [ "s71f6d301beba04", "s672f0b3115c8ff" ] }, { "age": 80.0, "six": 48.43416666666667, "five": 50.588, "tasks": { "blimp_filtered": 67.95, "supplement_filtered": 60.16, "ewok_filtered": 51.28, "entity_tracking": 21.26, "comps": 52.29, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 53.0 }, "checkpoint": "928e7d8f6614466e270b53335b4d79d840801af6db5ef5ddc62e6a5125f4ef3f", "precision": "FP32", "verification": "primary", "sources": [ "s98c473eb747e05", "s672f0b3115c8ff" ] }, { "age": 90.0, "six": 47.75833333333333, "five": 50.08, "tasks": { "blimp_filtered": 68.25, "supplement_filtered": 58.85, "ewok_filtered": 50.94, "entity_tracking": 20.32, "comps": 52.04, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 49.0 }, "checkpoint": "c0c481ffcdc22362e4fbaf0f0edb7bdb8a560c7182c61cbbacd142f7f6a12b13", "precision": "FP32", "verification": "primary", "sources": [ "s139eda7306ea7b", "s672f0b3115c8ff" ] }, { "age": 100.0, "six": 47.84166666666667, "five": 50.868, "tasks": { "blimp_filtered": 69.26, "supplement_filtered": 59.97, "ewok_filtered": 52.04, "entity_tracking": 20.85, "comps": 52.22, "global_piqa_parallel": 19.42, "global_piqa_nonparallel": 46.0 }, "checkpoint": "df841cd624e0c6d0f9524db5a03a53dbec29656afc130a5aa14c34b7163fcc2a", "precision": "FP32", "verification": "primary", "sources": [ "s97be831726da0b", "s672f0b3115c8ff" ] } ], "peak": { "five": { "score": 51.064, "ages": [ 70.0 ], "rank": 39 }, "six": { "score": 48.505, "ages": [ 70.0 ], "rank": 76 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 296, "id": "run/DC_TWO_PASS_A_D_OFF_POOLED_C_SEED20260907_100M_20261031", "original_name": "DC_TWO_PASS_A_D_OFF_POOLED_C_SEED20260907_100M_20261031", "title": "384幅・FFN1408・8層・DC2-pass A内Dなし", "topology": "単一出力", "recipe": "全群LR0.0004→0.00035→0.0003", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": "なし", "candidate": "未実現候補平均→共有C／処理差分→D", "reader": "D/C Key128→Value384", "received": "なし", "output": "単一出力", "reference": "8層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "なし", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "最終層", "reference_layers": [ [ 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "DC_TWO_MEMORY_SPHERICAL", "width": 384, "hidden": 1408, "layers": 8, "attention_heads": 12, "head_dim": 32, "output_head": "single_untied", "output_components": 1, "D_matrices": 7, "C_matrices": 1, "key_width": 128, "value_width": 384, "provisional_A_D": false, "candidate": "learned threshold pooled mean Key/Value into sharedC", "reader_width": 128, "attention_received_gate": false, "attention_qk_norm": true, "ffn_additive": false, "ffn_additive_bypass": false, "ffn_input_output_gates": false, "attention_distance": "none", "router_width": null, "output_state_layers": 1, "learn_rope_base": false, "learn_candidate_variance_scale": false }, "evaluation": { "tasks": "local6 with5 aggregate andallgenres", "precision": "FP32 TF32false", "recurrence": "exact pooled C truepreviousfinalprediction, canonical prefix; no two-pass inference approximation" }, "data": { "track": "strict-small", "packed_manifest_sha256": "611332e27d15ac01cd3f65cce4c09b5c291fd0150732ed2c134cc21abf4ba121", "words_per_pass": 10000000, "passes": 10, "boundary_policy": "zero start per packed row, attention and memory reset at document boundaries; does not carry history across separately packed windows" }, "tokenizer": { "vocab_size": 16384, "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8" }, "runtime": { "power_limit_watts": 260, "vram_budget_gib": 22, "precision": "BF16 linears/SDPA/FLA operands, FP32 master/residual/memory accumulator, FLA internal TF32; PyTorch TF32 false" }, "condition_conflicts": [], "profile": 167, "sources": [ "s2bff15c257bfe0" ], "points": [ { "age": 70.0, "six": 46.303333333333335, "five": 48.44, "tasks": { "blimp_filtered": 66.06, "supplement_filtered": 55.07, "ewok_filtered": 51.56, "entity_tracking": 18.8, "comps": 50.71, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 46.0 }, "checkpoint": "4806138ed835a03e2099e57256b1bc1b35ab190c343fc245a31935ef35052f75", "precision": "FP32", "verification": "primary", "sources": [ "s8703619047a5f2" ] }, { "age": 80.0, "six": 46.35166666666667, "five": 48.386, "tasks": { "blimp_filtered": 66.04, "supplement_filtered": 55.04, "ewok_filtered": 51.38, "entity_tracking": 18.42, "comps": 51.05, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 51.0 }, "checkpoint": "4e0dd8e62e79d997591d28e81f875e0e4bc6e564634d87a47286672d7108270b", "precision": "FP32", "verification": "primary", "sources": [ "sb9ebe12031ffc2" ] }, { "age": 90.0, "six": 46.035, "five": 48.126, "tasks": { "blimp_filtered": 65.7, "supplement_filtered": 54.47, "ewok_filtered": 51.17, "entity_tracking": 18.48, "comps": 50.81, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 43.0 }, "checkpoint": "9d138c6a82e76141f9852af84c3394470ec29295740a411ba4e0ee721fb7e743", "precision": "FP32", "verification": "primary", "sources": [ "seb98aa69648854" ] }, { "age": 100.0, "six": 45.495, "five": 47.864, "tasks": { "blimp_filtered": 65.5, "supplement_filtered": 53.19, "ewok_filtered": 50.74, "entity_tracking": 18.92, "comps": 50.97, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 44.0 }, "checkpoint": "08ea14b5a5bb5226c16e4ce19d955e6dd7ca8b413434b4c20684399aca92ec63", "precision": "FP32", "verification": "primary", "sources": [ "s21ab6c9031cd99" ] } ], "peak": { "five": { "score": 48.44, "ages": [ 70.0 ], "rank": 183 }, "six": { "score": 46.35166666666667, "ages": [ 80.0 ], "rank": 182 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 297, "id": "run/DC_TWO_PASS_A_D_ON_POOLED_C_SEED20260907_100M_20261031", "original_name": "DC_TWO_PASS_A_D_ON_POOLED_C_SEED20260907_100M_20261031", "title": "384幅・FFN1408・8層・DC2-pass A内Dあり", "topology": "単一出力", "recipe": "全群LR0.0004→0.00035→0.0003", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": "なし", "candidate": "未実現候補平均→共有C／処理差分→D", "reader": "D/C Key128→Value384", "received": "なし", "output": "単一出力", "reference": "8層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "なし", "candidate_var_learned": "なし", "mos": 1, "reference_mode": "最終層", "reference_layers": [ [ 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "DC_TWO_MEMORY_SPHERICAL", "width": 384, "hidden": 1408, "layers": 8, "attention_heads": 12, "head_dim": 32, "output_head": "single_untied", "output_components": 1, "D_matrices": 7, "C_matrices": 1, "key_width": 128, "value_width": 384, "provisional_A_D": true, "candidate": "learned threshold pooled mean Key/Value into sharedC", "reader_width": 128, "attention_received_gate": false, "attention_qk_norm": true, "ffn_additive": false, "ffn_additive_bypass": false, "ffn_input_output_gates": false, "attention_distance": "none", "router_width": null, "output_state_layers": 1, "learn_rope_base": false, "learn_candidate_variance_scale": false }, "evaluation": { "tasks": "local6 with5 aggregate andallgenres", "precision": "FP32 TF32false", "recurrence": "exact pooled C truepreviousfinalprediction, canonical prefix; no two-pass inference approximation" }, "data": { "track": "strict-small", "packed_manifest_sha256": "611332e27d15ac01cd3f65cce4c09b5c291fd0150732ed2c134cc21abf4ba121", "words_per_pass": 10000000, "passes": 10, "boundary_policy": "zero start per packed row, attention and memory reset at document boundaries; does not carry history across separately packed windows" }, "tokenizer": { "vocab_size": 16384, "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8" }, "runtime": { "power_limit_watts": 260, "vram_budget_gib": 22, "precision": "BF16 linears/SDPA/FLA operands, FP32 master/residual/memory accumulator, FLA internal TF32; PyTorch TF32 false" }, "condition_conflicts": [], "profile": 167, "sources": [ "sc561f8bdbced9e" ], "points": [ { "age": 70.0, "six": 46.98166666666667, "five": 48.568, "tasks": { "blimp_filtered": 67.64, "supplement_filtered": 54.72, "ewok_filtered": 49.55, "entity_tracking": 19.54, "comps": 51.39, "global_piqa_parallel": 30.1, "global_piqa_nonparallel": 48.0 }, "checkpoint": "8233904ce2296fe34e0d3da11c595ab1862688716dea4bfccf916e61184037c3", "precision": "FP32", "verification": "primary", "sources": [ "s8792b0954f31e2" ] }, { "age": 80.0, "six": 46.735, "five": 48.664, "tasks": { "blimp_filtered": 67.19, "supplement_filtered": 55.38, "ewok_filtered": 50.72, "entity_tracking": 19.19, "comps": 50.84, "global_piqa_parallel": 27.18, "global_piqa_nonparallel": 47.0 }, "checkpoint": "2f4bd41c65d7a260d9c33e69729adc34f361bdb08c9c7eec4c9e28538d4ab393", "precision": "FP32", "verification": "primary", "sources": [ "s9da15ed1fd19f0" ] }, { "age": 90.0, "six": 47.333333333333336, "five": 48.884, "tasks": { "blimp_filtered": 66.82, "supplement_filtered": 56.16, "ewok_filtered": 51.0, "entity_tracking": 19.62, "comps": 50.82, "global_piqa_parallel": 28.16, "global_piqa_nonparallel": 51.0 }, "checkpoint": "606708ececfe2df67e81bdcc0d6dead74918ba13f124deb59f0849a79eacde47", "precision": "FP32", "verification": "primary", "sources": [ "s2292bcdc7a06b3" ] }, { "age": 100.0, "six": 47.100833333333334, "five": 48.708, "tasks": { "blimp_filtered": 67.07, "supplement_filtered": 56.56, "ewok_filtered": 50.69, "entity_tracking": 18.44, "comps": 50.78, "global_piqa_parallel": 29.13, "global_piqa_nonparallel": 49.0 }, "checkpoint": "83c7c635c3547928f0fa565cc234f80a391d8fabccd2f2516412c0c66428ee2f", "precision": "FP32", "verification": "primary", "sources": [ "scb71a2b519beae" ] } ], "peak": { "five": { "score": 48.884, "ages": [ 90.0 ], "rank": 182 }, "six": { "score": 47.333333333333336, "ages": [ 90.0 ], "rank": 161 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 298, "id": "run/DC_D_ONLY_10M_20261031", "original_name": "DC_D_ONLY_10M_20261031", "title": "384幅・FFN1408・8層・Dのみ", "topology": "single untied linear softmax・枝別1層混合", "recipe": "行単位・可変更新・LR400固定", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": "なし", "candidate": "候補記憶なし／処理差分→D", "reader": "Dのみ Key128→Value384", "received": "なし", "output": "single untied linear softmax", "reference": "8層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "なし", "candidate_var_learned": "未確認", "mos": null, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "DC_D_ONLY_SPHERICAL", "width": 384, "hidden": 1408, "layers": 8, "attention_heads": 12, "head_dim": 32, "router_width": 0, "candidate_reader_enabled": false, "reader_width": 0, "attention_received_gate": false, "output_head": "single untied linear softmax", "output_branch_layers": [ [ 8 ] ], "ffn_additive_bypass": false, "attention_qk_norm": true, "attention_distance": "none", "ffn_input_output_gates": false, "learn_rope_base": false, "rope_base": 10000, "D_memories": 7, "D_key": 128, "D_value": 384, "C_memories": 0, "seed": 20260907 }, "evaluation": { "tasks": [ [ "blimp", "blimp_filtered", 59875 ], [ "blimp", "supplement_filtered", 5218 ], [ "ewok", "ewok_filtered", 7618 ], [ "entity_tracking", "entity_tracking", 6780 ], [ "comps", "comps", 91028 ], [ "global_piqa_parallel", "global_piqa_parallel", 103 ], [ "global_piqa_nonparallel", "global_piqa_nonparallel", 100 ] ], "precision": "FP32 TF32false; exact D recurrence; canonical prefix and phrase", "official_prepared_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28" }, "data": { "source_main": "source:s8b3653d38e3b18", "source_mini": "source:scdc7066e86821c", "epoch": 0, "manifest_sha256": "611332e27d15ac01cd3f65cce4c09b5c291fd0150732ed2c134cc21abf4ba121", "target_positions_per_update": 4096, "actual_updates": 3489, "microbatch_positions": 4096, "max_documents": 256, "min_bucket": 16, "labels": "original frozen physical-row-end omissions preserved", "window": "same parent row-EOS whole-document windows; data/order/global update unchanged", "state": "full gradient through true document; reset only at document boundaries", "epochs_remaining": [ 1, 2, 3, 4, 5, 6, 7, 8, 9 ] }, "tokenizer": {}, "runtime": { "D": "FP32 simultaneous rank-two recompute scan", "attention": "BF16 SDPA, same QK normalization and RoPE10000", "state": "FP32 spherical", "tf32": false, "deterministic": true, "compile": "expanded D inputs + fused D-read/attention/FFN body, emulate_precision_casts", "fuse_body": true, "api_call_length": 0, "api_call_length_meaning": "whole padded document, no C publication splitting", "power_W": 260, "allocator_cap_GiB": 21.4, "vram_limit_GiB": 22 }, "condition_conflicts": [], "profile": 35, "sources": [ "s640a2daa6ce0ae" ], "points": [ { "age": 70.0, "six": 47.594166666666666, "five": 49.286, "tasks": { "blimp_filtered": 66.76, "supplement_filtered": 58.51, "ewok_filtered": 51.34, "entity_tracking": 19.17, "comps": 50.65, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 54.0 }, "checkpoint": "daae1c8079f94b04b459b7230037e4eee8a1451834055ef7dd24c29c9511c1d1", "precision": "FP32", "verification": "primary", "sources": [ "s25f86318b0e1cd" ] }, { "age": 80.0, "six": 47.54666666666667, "five": 50.026, "tasks": { "blimp_filtered": 67.52, "supplement_filtered": 61.65, "ewok_filtered": 50.85, "entity_tracking": 18.93, "comps": 51.18, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 47.0 }, "checkpoint": "a6226240480bec96a823740bcd40736434367f75761075326903a572eba29bd9", "precision": "FP32", "verification": "primary", "sources": [ "sbcff84752688a7" ] }, { "age": 90.0, "six": 47.38666666666666, "five": 49.44, "tasks": { "blimp_filtered": 66.27, "supplement_filtered": 59.48, "ewok_filtered": 51.55, "entity_tracking": 18.98, "comps": 50.92, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 49.0 }, "checkpoint": "f7d01307b78386ed7da7861b3219b6de7a95e19e94eeb050da723766ad76be5d", "precision": "FP32", "verification": "primary", "sources": [ "s12214a79bd4930" ] }, { "age": 100.0, "six": 47.38916666666667, "five": 48.96, "tasks": { "blimp_filtered": 66.42, "supplement_filtered": 56.94, "ewok_filtered": 52.65, "entity_tracking": 18.21, "comps": 50.58, "global_piqa_parallel": 31.07, "global_piqa_nonparallel": 48.0 }, "checkpoint": "92eb29760d603e3c19add51e54215fec362eaa28b4cf834fabb3764b1cd2049f", "precision": "FP32", "verification": "primary", "sources": [ "seb0dfb3380abed" ] } ], "peak": { "five": { "score": 50.026, "ages": [ 80.0 ], "rank": 149 }, "six": { "score": 47.594166666666666, "ages": [ 70.0 ], "rank": 141 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 299, "id": "run/DC_D_ONLY_GROUPED4096_100M_20261101", "original_name": "DC_D_ONLY_GROUPED4096_100M_20261101", "title": "384幅・FFN1408・8層・Dのみ", "topology": "単一出力", "recipe": "隣接行グループ・4096固定更新・LR400→350→300", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": "なし", "candidate": "候補記憶なし／処理差分→D", "reader": "Dのみ Key128→Value384", "received": "なし", "output": "単一出力", "reference": "8層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "なし", "candidate_var_learned": "未確認", "mos": 1, "reference_mode": "最終層", "reference_layers": [ [ 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "DC_D_ONLY_SPHERICAL", "width": 384, "hidden": 1408, "layers": 8, "attention_heads": 12, "head_dim": 32, "router_width": 0, "candidate_reader_enabled": false, "reader_width": 0, "attention_received_gate": false, "output_head": "single untied linear softmax", "ffn_additive_bypass": false, "attention_qk_norm": true, "attention_distance": "none", "ffn_input_output_gates": false, "learn_rope_base": false, "rope_base": 10000, "D_memories": 7, "D_key": 128, "D_value": 384, "C_memories": 0, "seed": 20260907, "output_components": 1, "output_state_layers": 1 }, "evaluation": { "tasks": [ [ "blimp", "blimp_filtered", 59875 ], [ "blimp", "supplement_filtered", 5218 ], [ "ewok", "ewok_filtered", 7618 ], [ "entity_tracking", "entity_tracking", 6780 ], [ "comps", "comps", 91028 ], [ "global_piqa_parallel", "global_piqa_parallel", 103 ], [ "global_piqa_nonparallel", "global_piqa_nonparallel", 100 ] ], "precision": "FP32 TF32false; exact D recurrence; canonical prefix and phrase", "official_prepared_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28" }, "data": { "track": "BabyLM2026-English-Strict-Small; owner-directed local causal experiment", "raw_unique_words": 10000000, "total_input_word_exposure": 100000000, "source_main": "source:sc12637a046848a", "source_mini": "source:s9bf670d77165e0", "data_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "tokenizer_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "logical_lengths": [ 128, 256, 512, 128, 256, 512, 128, 256, 512, 512 ], "context": "adjacent-line groups as #277; reset also at declared logical context window; not true document IDs", "batch_positions": 4096, "last_batch": "one partial update at each epoch end", "loss": "per-update mean across all supervised positions; mask logical-window/group crossings", "schedule": { "base_lrs": [ 0.0004, 0.00035, 0.0003 ], "milestones_words": [ 0, 30000000, 60000000 ], "warmup_words": 1600000, "dropout": [ 0.0, 0.02, 0.05 ], "beta1": 0.9, "beta2": 0.9872585449014338, "eps": 1e-08, "clip": 1.0, "wd_matrix_lexical": 0.1, "wd_control": 0.0, "schedule_lookup": "words consumed BEFORE the next update; warmup=(words+batch_words)/1600000 capped1; same #277 convention" }, "runtime": { "d_kernel": "recompute", "fuse_body": true, "precision": "BF16 body, FP32 D/state/geometry", "tf32": false, "microbatch_positions": 8192, "min_bucket": 16, "compile_recompile_limit": 64, "deterministic": true, "power_W": 260, "vram_cap_GiB": 22 } }, "tokenizer": {}, "runtime": { "d_kernel": "recompute", "fuse_body": true, "precision": "BF16 body, FP32 D/state/geometry", "tf32": false, "microbatch_positions": 8192, "min_bucket": 16, "compile_recompile_limit": 64, "deterministic": true, "power_W": 260, "vram_cap_GiB": 22 }, "condition_conflicts": [], "profile": 150, "sources": [ "s44ce4b727981f4" ], "points": [ { "age": 70.0, "six": 47.11666666666667, "five": 49.892, "tasks": { "blimp_filtered": 68.74, "supplement_filtered": 57.97, "ewok_filtered": 50.49, "entity_tracking": 20.69, "comps": 51.57, "global_piqa_parallel": 17.48, "global_piqa_nonparallel": 49.0 }, "checkpoint": "0411727b33dfee82b84e8a6a19e50c039f2f3b73230653b9352a63455f7a1ce6", "precision": "FP32", "verification": "primary", "sources": [ "s3cc6143e14c09c" ] }, { "age": 80.0, "six": 47.44, "five": 50.272, "tasks": { "blimp_filtered": 67.99, "supplement_filtered": 58.9, "ewok_filtered": 51.45, "entity_tracking": 20.73, "comps": 52.29, "global_piqa_parallel": 14.56, "global_piqa_nonparallel": 52.0 }, "checkpoint": "a153196d2b82484a86ac2ddd34f95165853719409fa7e5d76c176450608465ac", "precision": "FP32", "verification": "primary", "sources": [ "s8e5a583cdf2998" ] }, { "age": 90.0, "six": 47.285833333333336, "five": 49.804, "tasks": { "blimp_filtered": 67.27, "supplement_filtered": 59.29, "ewok_filtered": 51.49, "entity_tracking": 19.58, "comps": 51.39, "global_piqa_parallel": 20.39, "global_piqa_nonparallel": 49.0 }, "checkpoint": "5e3b668a26a2acb87cdfeb9c38b06dc249754bf089d3f1c7a889f3aa2128d116", "precision": "FP32", "verification": "primary", "sources": [ "sf0e9dfd00032dc" ] }, { "age": 100.0, "six": 46.6675, "five": 49.856, "tasks": { "blimp_filtered": 67.33, "supplement_filtered": 58.93, "ewok_filtered": 50.91, "entity_tracking": 20.26, "comps": 51.85, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 43.0 }, "checkpoint": "0f7672830dfb7fa4517032546f34d5a9159dc40d3b23b32a520a896f5a90ba7d", "precision": "FP32", "verification": "primary", "sources": [ "sc56ac74d6fdd07" ] } ], "peak": { "five": { "score": 50.272, "ages": [ 80.0 ], "rank": 126 }, "six": { "score": 47.44, "ages": [ 80.0 ], "rank": 147 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 300, "id": "run/DC_D_READ_GATE_100M_20261101", "original_name": "DC_D_READ_GATE_100M_20261101", "title": "384幅・FFN1408・8層・Dのみ＋入力依存読出しゲート", "topology": "単一出力", "recipe": "隣接行グループ・4096固定更新・LR400→350→300", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": "なし", "candidate": "候補記憶なし／処理差分→D", "reader": "Dのみ Key128→Value384／入力依存読出しゲート", "received": "なし", "output": "単一出力", "reference": "8層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "なし", "candidate_var_learned": "未確認", "mos": 1, "reference_mode": "最終層", "reference_layers": [ [ 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "DC_D_ONLY_SPHERICAL_READ_ADOPTION", "width": 384, "hidden": 1408, "layers": 8, "attention_heads": 12, "head_dim": 32, "router_width": 0, "candidate_reader_enabled": false, "reader_width": 0, "attention_received_gate": false, "output_head": "single untied linear softmax", "ffn_additive_bypass": false, "attention_qk_norm": true, "attention_distance": "none", "ffn_input_output_gates": false, "learn_rope_base": false, "rope_base": 10000, "D_memories": 7, "D_key": 128, "D_value": 384, "C_memories": 0, "seed": 20260907, "output_components": 1, "output_state_layers": 1, "D_read_adoption": { "gate": "2sigmoid(wh.h + wc.correction + b)", "shape": "one scalar per position per D layer", "init": "all zero ->1", "input_detach": false, "new_parameters": 5383 } }, "evaluation": { "tasks": [ [ "blimp", "blimp_filtered", 59875 ], [ "blimp", "supplement_filtered", 5218 ], [ "ewok", "ewok_filtered", 7618 ], [ "entity_tracking", "entity_tracking", 6780 ], [ "comps", "comps", 91028 ], [ "global_piqa_parallel", "global_piqa_parallel", 103 ], [ "global_piqa_nonparallel", "global_piqa_nonparallel", 100 ] ], "precision": "FP32 TF32false; exact D recurrence; canonical prefix and phrase", "official_prepared_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28" }, "data": { "track": "BabyLM2026-English-Strict-Small; owner-directed local causal experiment", "raw_unique_words": 10000000, "total_input_word_exposure": 100000000, "source_main": "source:sc12637a046848a", "source_mini": "source:s9bf670d77165e0", "data_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "tokenizer_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "logical_lengths": [ 128, 256, 512, 128, 256, 512, 128, 256, 512, 512 ], "context": "adjacent-line groups as #277; reset also at declared logical context window; not true document IDs", "batch_positions": 4096, "last_batch": "one partial update at each epoch end", "loss": "per-update mean across all supervised positions; mask logical-window/group crossings", "schedule": { "base_lrs": [ 0.0004, 0.00035, 0.0003 ], "milestones_words": [ 0, 30000000, 60000000 ], "warmup_words": 1600000, "dropout": [ 0.0, 0.02, 0.05 ], "beta1": 0.9, "beta2": 0.9872585449014338, "eps": 1e-08, "clip": 1.0, "wd_matrix_lexical": 0.1, "wd_control": 0.0, "schedule_lookup": "words consumed BEFORE the next update; warmup=(words+batch_words)/1600000 capped1; same #277 convention" }, "runtime": { "d_kernel": "recompute", "fuse_body": true, "precision": "BF16 body, FP32 D/state/geometry", "tf32": false, "microbatch_positions": 8192, "min_bucket": 16, "compile_recompile_limit": 64, "deterministic": true, "power_W": 260, "vram_cap_GiB": 22 } }, "tokenizer": {}, "runtime": { "d_kernel": "recompute", "fuse_body": true, "precision": "BF16 body, FP32 D/state/geometry", "tf32": false, "microbatch_positions": 8192, "min_bucket": 16, "compile_recompile_limit": 64, "deterministic": true, "power_W": 260, "vram_cap_GiB": 22 }, "condition_conflicts": [], "profile": 151, "sources": [ "se7d3b9533eafcb" ], "points": [ { "age": 70.0, "six": 47.0575, "five": 49.736, "tasks": { "blimp_filtered": 68.34, "supplement_filtered": 56.66, "ewok_filtered": 51.08, "entity_tracking": 20.32, "comps": 52.28, "global_piqa_parallel": 22.33, "global_piqa_nonparallel": 45.0 }, "checkpoint": "a4c04f5dbc70a725fa844d26d228a3ecbe2582b84ea66031971d34e6917d3d7d", "precision": "FP32", "verification": "primary", "sources": [ "s9e5ad1adb520ab" ] }, { "age": 80.0, "six": 47.3125, "five": 49.93, "tasks": { "blimp_filtered": 67.33, "supplement_filtered": 58.43, "ewok_filtered": 50.78, "entity_tracking": 20.74, "comps": 52.37, "global_piqa_parallel": 18.45, "global_piqa_nonparallel": 50.0 }, "checkpoint": "abed7b1f2c3152aedf3130c24cacea495ae3fae9c9705536b33af14df202c327", "precision": "FP32", "verification": "primary", "sources": [ "s2423d123c0d94b" ] }, { "age": 90.0, "six": 46.93833333333333, "five": 49.59, "tasks": { "blimp_filtered": 66.67, "supplement_filtered": 58.15, "ewok_filtered": 50.83, "entity_tracking": 20.07, "comps": 52.23, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 46.0 }, "checkpoint": "74c88796adaec845486d415d595178e30be868277503d9ff809b52043600e3cd", "precision": "FP32", "verification": "primary", "sources": [ "s949dc5a243b7ca" ] }, { "age": 100.0, "six": 46.9825, "five": 49.552, "tasks": { "blimp_filtered": 67.33, "supplement_filtered": 56.44, "ewok_filtered": 49.97, "entity_tracking": 21.73, "comps": 52.29, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 44.0 }, "checkpoint": "664c0ae7b8d37c9b8ca1e4a04af3fee5baadaee7d09b90916adad4140240732e", "precision": "FP32", "verification": "primary", "sources": [ "s0dd10bdda124b2" ] } ], "peak": { "five": { "score": 49.93, "ages": [ 80.0 ], "rank": 155 }, "six": { "score": 47.3125, "ages": [ 80.0 ], "rank": 162 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 301, "id": "run/DC_DDN_GROUPED4096_100M_20261101", "original_name": "DC_DDN_GROUPED4096_100M_20261101", "title": "384幅・FFN1408・8層・Dのみ・Key/Value両側修正", "topology": "単一出力", "recipe": "隣接行グループ・4096固定更新・LR400→350→300", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 20260907, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": "なし", "candidate": "候補記憶なし／処理差分→D", "reader": "Dのみ Key128→Value384／Key・Value両側修正", "received": "なし", "output": "単一出力", "reference": "8層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "なし", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "なし", "candidate_var_learned": "未確認", "mos": 1, "reference_mode": "最終層", "reference_layers": [ [ 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [], "state": "採点済み", "architecture": { "id": "DC_D_ONLY_SPHERICAL_DDN", "width": 384, "hidden": 1408, "layers": 8, "attention_heads": 12, "head_dim": 32, "router_width": 0, "candidate_reader_enabled": false, "reader_width": 0, "attention_received_gate": false, "output_head": "single untied linear softmax", "ffn_additive_bypass": false, "attention_qk_norm": true, "attention_distance": "none", "ffn_input_output_gates": false, "learn_rope_base": false, "rope_base": 10000, "D_memories": 7, "D_key": 128, "D_value": 384, "C_memories": 0, "seed": 20260907, "output_components": 1, "output_state_layers": 1, "D_dual_update": true, "implementation_id": "DC-DDN-simultaneous-two-write-v1", "ffn": 1408, "D_shape": [ 128, 384 ], "D_count": 7, "dual_selector_initial": 0.5, "dual_gamma_initial_total": 0.05, "value_normalization_epsilon": 1e-06 }, "evaluation": { "tasks": [ [ "blimp", "blimp_filtered", 59875 ], [ "blimp", "supplement_filtered", 5218 ], [ "ewok", "ewok_filtered", 7618 ], [ "entity_tracking", "entity_tracking", 6780 ], [ "comps", "comps", 91028 ], [ "global_piqa_parallel", "global_piqa_parallel", 103 ], [ "global_piqa_nonparallel", "global_piqa_nonparallel", 100 ] ], "precision": "FP32 TF32false; exact D recurrence; canonical prefix and phrase", "official_prepared_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28" }, "data": { "track": "BabyLM2026-English-Strict-Small; owner-directed local causal experiment", "raw_unique_words": 10000000, "total_input_word_exposure": 100000000, "source_main": "source:sc12637a046848a", "source_mini": "source:s9bf670d77165e0", "data_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "tokenizer_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "logical_lengths": [ 128, 256, 512, 128, 256, 512, 128, 256, 512, 512 ], "context": "adjacent-line groups as #277; reset also at declared logical context window; not true document IDs", "batch_positions": 4096, "last_batch": "one partial update at each epoch end", "loss": "per-update mean across all supervised positions; mask logical-window/group crossings", "schedule": { "base_lrs": [ 0.0004, 0.00035, 0.0003 ], "milestones_words": [ 0, 30000000, 60000000 ], "warmup_words": 1600000, "dropout": [ 0.0, 0.02, 0.05 ], "beta1": 0.9, "beta2": 0.9872585449014338, "eps": 1e-08, "clip": 1.0, "wd_matrix_lexical": 0.1, "wd_control": 0.0, "schedule_lookup": "words consumed BEFORE the next update; warmup=(words+batch_words)/1600000 capped1; same #277 convention" }, "runtime": { "d_kernel": "parallel two-sided block-WY, chunk32, sparse2x2, FP32 storage/accum with TF32x3 products in training, IEEE evaluation", "fuse_body": true, "precision": "BF16 body, FP32 D/state/geometry", "tf32": false, "microbatch_positions": 8192, "min_bucket": 64, "compile_recompile_limit": 64, "deterministic": true, "power_W": 260, "vram_cap_GiB": 22, "d_recurrence": "DDN simultaneous two-write", "dual_gamma_initial_total": 0.05, "selector_initial": 0.5, "d_chunk": 32, "d_inverse": "finite nilpotent doubling; no algebraic truncation", "d_matmul": "FP32 storage/output/accumulation with TF32x3 tensor-core products in training D chunk; FP32 IEEE in evaluation", "d_cuda_graphs": true, "d_block_sparse_exact": true } }, "tokenizer": {}, "runtime": { "d_kernel": "parallel two-sided block-WY, chunk32, sparse2x2, FP32 storage/accum with TF32x3 products in training, IEEE evaluation", "fuse_body": true, "precision": "BF16 body, FP32 D/state/geometry", "tf32": false, "microbatch_positions": 8192, "min_bucket": 64, "compile_recompile_limit": 64, "deterministic": true, "power_W": 260, "vram_cap_GiB": 22, "d_recurrence": "DDN simultaneous two-write", "dual_gamma_initial_total": 0.05, "selector_initial": 0.5, "d_chunk": 32, "d_inverse": "finite nilpotent doubling; no algebraic truncation", "d_matmul": "FP32 storage/output/accumulation with TF32x3 tensor-core products in training D chunk; FP32 IEEE in evaluation", "d_cuda_graphs": true, "d_block_sparse_exact": true }, "condition_conflicts": [], "profile": 149, "sources": [ "s83e43d56cf985d" ], "points": [ { "age": 70.0, "six": 47.7225, "five": 50.24, "tasks": { "blimp_filtered": 68.15, "supplement_filtered": 58.53, "ewok_filtered": 51.36, "entity_tracking": 21.25, "comps": 51.91, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 46.0 }, "checkpoint": "6b53e622c788af53309253efcac49d1b7ef11921d2475c705ed28a5d7bd49f14", "precision": "FP32", "verification": "primary", "sources": [ "sccec2670f114f7" ] }, { "age": 80.0, "six": 47.32833333333333, "five": 50.258, "tasks": { "blimp_filtered": 67.43, "supplement_filtered": 59.94, "ewok_filtered": 51.45, "entity_tracking": 20.7, "comps": 51.77, "global_piqa_parallel": 21.36, "global_piqa_nonparallel": 44.0 }, "checkpoint": "73f345f5bfcc326dc43cea16fa2690d21a09feafcde69c64ad1e3bee97021ee8", "precision": "FP32", "verification": "primary", "sources": [ "sd885cab7debc4c" ] }, { "age": 90.0, "six": 46.97666666666667, "five": 49.948, "tasks": { "blimp_filtered": 67.87, "supplement_filtered": 59.27, "ewok_filtered": 51.33, "entity_tracking": 19.68, "comps": 51.59, "global_piqa_parallel": 25.24, "global_piqa_nonparallel": 39.0 }, "checkpoint": "7f3cc72d6e8f32a8a2b3a36fcb7e9635f58e194311575ce4343c923b89d5da33", "precision": "FP32", "verification": "primary", "sources": [ "sdc98b768d11047" ] }, { "age": 100.0, "six": 47.093333333333334, "five": 49.882, "tasks": { "blimp_filtered": 68.22, "supplement_filtered": 56.61, "ewok_filtered": 51.22, "entity_tracking": 21.52, "comps": 51.84, "global_piqa_parallel": 23.3, "global_piqa_nonparallel": 43.0 }, "checkpoint": "fdd76947a49180418562aeea7009f8eba6da9fbac789c0172fdb3926fba1a18b", "precision": "FP32", "verification": "primary", "sources": [ "s79ab5acda25600" ] } ], "peak": { "five": { "score": 50.258, "ages": [ 80.0 ], "rank": 130 }, "six": { "score": 47.7225, "ages": [ 70.0 ], "rank": 135 } }, "clear_points": [], "clear_peak": null, "present": true }, { "number": 302, "id": "run/IPRLM384_MODEL277_SEED137_80M_20261101", "original_name": "IPRLM384_MODEL277_SEED137_80M_20261101", "title": "384幅・FFN1408・8層", "topology": "MoS3・枝別1/2/3層混合", "recipe": "主要0.00048／Embedding0.00084", "cohort": "baby", "cohort_label": "Baby同系列", "seed": 137, "width": 384, "ffn": 1408, "depth": 8, "parts": { "width": 384, "ffn": 1408, "depth": 8, "attention": "12×32", "router": 32, "candidate": "候補の平均・分散を帰還", "reader": 48, "received": "あり", "output": "MoS3", "reference": "8層 / 7・8層 / 6・7・8層", "ffn_add": "なし", "qk": "あり", "distance": "なし", "ffn_gates": "未確認", "ffn_g_norm": "未確認", "bypass_condition": "未確認", "mixture_gate": "未確認", "rope_learned": "あり", "candidate_var_learned": "あり", "mos": 3, "reference_mode": "枝別に混合", "reference_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_ratio": 3.6666666666666665 }, "main_lr": [ 0.00048, 0.00041999999999999996, 0.00035999999999999997 ], "state": "採点済み", "architecture": { "id": "MODEL3_P", "width": 384, "layers": 8, "hidden": 1408, "attention_heads": 12, "head_dim": 32, "alpha_init": 0.2, "tau_init": 1.0, "variance_scale_init": 0.1, "gate_bias_init": 1.0, "router_width": 32, "passes": 2, "geometry": "raw", "knob_lr_ratio": 0.3, "dropout_init": 0.0, "skip_bos_transition": true, "reader_width": 48, "attention_distance": "none", "learn_rope_base": true, "rope_base_init": 10000.0, "learn_candidate_variance_scale": true, "candidate_variance_scale_init": 0.1, "history_norm": "separate", "seed": 137, "attention_width": 384, "width_scaling": "fan_in_ratio", "reference_width": 256, "reference_hidden": 896, "attention_qk_norm": true, "qk_norm_eps": 1e-06, "output_head": "multi_input_mos3_branch123", "attention_received_gate": true, "output_components": 3, "output_state_layers": 3, "reference_attention_width": 256, "output_branch_layers": [ [ 8 ], [ 7, 8 ], [ 6, 7, 8 ] ], "ffn_scaling_anchor_hidden": 1120, "parent_id": "VISTARC_CANDIDATE_READER", "plastic_owners": 104, "plastic_rank": 16, "plastic_state": "owner-local F", "plastic_target": "arriving real-vocabulary error", "plastic_readout": "Query, received gate and FFN native input", "initialization_depth_anchor": 9, "ffn_additive_bypass": false, "candidate_handling": "候補の平均・分散を帰還" }, "evaluation": { "suite": "local six-axis", "mode": "exact recurrent canonical prefix", "precision": "FP32", "tf32": false, "frozen_eval_runtime_sha256": "a8280b292806610a2e7d5e3c9d21dd42583776bda888159938666b5202177d25", "prepared_manifest_sha256": "517e9ebccc43edca357a7fbeebe4a1b0a2b5d5e456363300384753e7b0fbeb28", "benchmark_words": [ 80000000 ], "timing": "after80M", "official_commit": "6f825c291e2c4c78ad33b1935fd64d45f52642dc", "comparison": "#1/#81/#26 and #277/#294, matched80M three-seed means and full inventory", "supervised7": false }, "data": { "official_dataset": "BabyLM-community/BabyLM-2026-Strict-Small", "official_revision": "c92ab16b4f08858304b0815706065b3354d8fc0a", "documents": 120011, "raw_sources": [ "bnc_spoken.train.txt", "childes.train.txt", "gutenberg.train.txt", "open_subtitles.train.txt", "simple_wiki.train.txt", "switchboard.train.txt" ], "packed_path": "data/cycle", "packed_manifest_sha256": "c532445ffee9a597d7d420e65f37278d6d39a6522e890b8754e43956d7c19b42", "words_per_pass": 10000000, "stream_readback": "PASS_EXACT_GROUP_ORDER_NO_OMISSIONS", "boundary_policy": "group-isolated attention and feedback; original adjacent lines inside groups share context", "segment_id_policy": "on-disk zero-based real segments; loader maps valid positions to one-based segments", "segment_validity": "column < lengths; 0 reserved for padding after mapping", "original_lines": 1104106, "grouping_audit_sha256": "e064017f84267130369a07405b679d03a2d2656c74835d53a2d5b577f63aec10", "grouping": "Adjacent original lines, target128 content tokens, max256 except long single lines, max16 lines; reset at source or visible marker", "boundary_limit": "Official files omit document IDs; unmarked conversation/book boundaries cannot be recovered", "run_manifest": "CPU_DATA_ACCEPTANCE.json" }, "tokenizer": { "path": "data/tokenizer", "type": "ByteLevel BPE", "tokenizer_json_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", "vocab_size": 16384, "unk_token_id": 0, "bos_token_id": 1, "eos_token_id": 2, "pad_token_id": 3, "mask_token_id": 4 }, "runtime": { "precision": "BF16 training", "variant": "all", "scan_chunk": 32, "F_update": "each arrival", "compiled_source": "source", "GPU": "RTX3090", "max_vram_gib": 22, "power_limit_watts": 260, "vram_budget_gib": 22, "recompile_limit": 32, "training_update": "all+outer_cuda_graph; compiled forward/backward only, AdamW/clip and metadata validation outside capture", "cuda_graph_origin": "#276/extensions/steady_speed_20261028/graph_update.py; revalidated at8layers before fresh training" }, "condition_conflicts": [], "profile": 70, "sources": [ "s467308bf40589b" ], "points": [ { "age": 80.0, "six": 49.0175, "five": 50.894, "tasks": { "blimp_filtered": 68.45, "supplement_filtered": 59.68, "ewok_filtered": 54.1, "entity_tracking": 19.85, "comps": 52.39, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 55.0 }, "checkpoint": "437b5f5c4445840dd7ca692d3e195d1661d2ffbb648efc1d7e4dcde3274d1f8b", "precision": "FP32", "verification": "primary", "sources": [ "se971cfa3885036", "s98eb35aa1b8f8d" ] } ], "peak": { "five": { "score": 50.894, "ages": [ 80.0 ], "rank": 59 }, "six": { "score": 49.0175, "ages": [ 80.0 ], "rank": 31 } }, "clear_points": [ { "number": 302, "age": 80, "seed": 137, "checkpoint_sha256": "437b5f5c4445840dd7ca692d3e195d1661d2ffbb648efc1d7e4dcde3274d1f8b", "old_six": 49.0175, "old_five": 50.894, "old_tasks": { "blimp_filtered": 68.45, "supplement_filtered": 59.68, "ewok_filtered": 54.1, "entity_tracking": 19.85, "comps": 52.39, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 55.0 }, "correct": 11987, "total": 17274, "accuracy": 69.39330786152598, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6759, "total": 8392, "ties": 0 }, "comps": { "correct": 3827, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 15, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 987, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 80, "total": 114, "ties": 0 }, "boolq": { "correct": 15, "total": 53, "ties": 0 }, "mnli": { "correct": 25, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 243.82449291300145, "checkpoint_rank": 20, "rank": 16 } ], "clear_peak": { "number": 302, "age": 80, "seed": 137, "checkpoint_sha256": "437b5f5c4445840dd7ca692d3e195d1661d2ffbb648efc1d7e4dcde3274d1f8b", "old_six": 49.0175, "old_five": 50.894, "old_tasks": { "blimp_filtered": 68.45, "supplement_filtered": 59.68, "ewok_filtered": 54.1, "entity_tracking": 19.85, "comps": 52.39, "global_piqa_parallel": 24.27, "global_piqa_nonparallel": 55.0 }, "correct": 11987, "total": 17274, "accuracy": 69.39330786152598, "ties": 0, "tasks": { "blimp_filtered": { "correct": 6759, "total": 8392, "ties": 0 }, "comps": { "correct": 3827, "total": 6299, "ties": 0 }, "entity_tracking": { "correct": 2, "total": 10, "ties": 0 }, "ewok_filtered": { "correct": 15, "total": 30, "ties": 0 }, "global_piqa_nonparallel": { "correct": 1, "total": 1, "ties": 0 }, "global_piqa_parallel": { "correct": 0, "total": 1, "ties": 0 }, "supplement_filtered": { "correct": 987, "total": 1106, "ties": 0 }, "babi_qa01": { "correct": 0, "total": 6, "ties": 0 }, "babi_qa11": { "correct": 1, "total": 11, "ties": 0 }, "babi_qa12": { "correct": 1, "total": 5, "ties": 0 }, "babi_qa13": { "correct": 1, "total": 3, "ties": 0 }, "babi_qa14": { "correct": 0, "total": 1, "ties": 0 }, "proofwriter_depth0": { "correct": 3, "total": 6, "ties": 0 }, "proofwriter_depth1": { "correct": 3, "total": 5, "ties": 0 }, "proofwriter_depth2": { "correct": 4, "total": 7, "ties": 0 }, "proofwriter_depth3": { "correct": 1, "total": 2, "ties": 0 }, "proofwriter_depth5": { "correct": 1, "total": 2, "ties": 0 }, "piqa1838": { "correct": 80, "total": 114, "ties": 0 }, "boolq": { "correct": 15, "total": 53, "ties": 0 }, "mnli": { "correct": 25, "total": 203, "ties": 0 }, "mrpc": { "correct": 9, "total": 10, "ties": 0 }, "multirc": { "correct": 40, "total": 52, "ties": 0 }, "qqp": { "correct": 211, "total": 947, "ties": 0 }, "rte": { "correct": 1, "total": 8, "ties": 0 } }, "status": "complete", "gpu_seconds": 243.82449291300145, "checkpoint_rank": 20, "rank": 16 }, "present": true } ]