| { |
| "gate_passed": true, |
| "phase2_historical_bpb": 1.4102684308437439, |
| "phase3_historical_bpb": 1.0653798981287268, |
| "required_max_historical_bpb": 1.3397550093015567, |
| "historical_bpb_improvement_rate_over_phase2": 0.24455480416, |
| "phase3_modern_control_bpb": 1.8756087153420742, |
| "historical_lower_than_modern": true, |
| "automatic_rubric_pass_rate": 0.9333333333333333, |
| "manual_rubric_pass_rate": 0.9, |
| "manual_review": { |
| "reviewed_generations": 5, |
| "checks": 10, |
| "passed_checks": 9, |
| "empty_outputs": 0, |
| "mojibake_outputs": 0, |
| "unbounded_repetition_outputs": 0, |
| "historical_style_outputs": 5, |
| "severely_broken_outputs": 0, |
| "noted_issue": "The 人工知能 probe is flagged no_mojibake=false by the automatic check, but the flagged characters are period old-kana renderings (存すへき=存すべき, 得へし), not true mojibake; the 文明開化 probe has elevated but bounded repetition (0.16)." |
| }, |
| "training": { |
| "iterations": 11445, |
| "processed_tokens": 3000238080, |
| "epochs_over_train": 2.72, |
| "best_validation_loss": 4.149158000946045, |
| "elapsed_hours": 18.29, |
| "estimated_cost_usd": 18.11, |
| "peak_vram_bytes": 44345157632, |
| "gpu_total_bytes": 47803596800 |
| }, |
| "final_checkpoint_sha256": "1f2459a1b5b5c75c6e7c07251b877f6c804500753bb0121b13351933b8a07b03", |
| "tokenizer_sha256": "2c014c2f018dd70ab48d66b59e20c9afc12c5a0dd7b5bbdd48d7b849c4bf1d42", |
| "vocab_size": 16384, |
| "parameter_count": 345035776, |
| "review_date": "2026-07-17", |
| "notes": [ |
| "Phase-3 historical test BPB 1.0654 beats the phase-2 gate (<= 1.3398, i.e. 5% better than phase-2's 1.4103) by a 24.5% improvement and is well below the modern control (1.8756).", |
| "Single-pass full run cleared the gate; no continued-training round was needed (unlike phase-2's 1,600-step shortfall)." |
| ] |
| } |
|
|
