| { |
| "session": "v14-v15-long-context-adjudication-and-teacher-rendering", |
| "model": "Qwen3-0.6B Alea", |
| "base_checkpoint": "v12_rl_cal", |
| "public_tier_metric_note": "These are reproducible local accuracies on the frozen public JevBench JSONL tiers, computed over all rows (choice/score argmax; noul threshold 0.5). This is not the sealed-set v1.4 composite. These direct re-evaluations do not match some older session-reported JevBench values, so treat historical and current scores as non-comparable until the previous runner/predictions are recovered.", |
| "public_tier_results_percent": { |
| "v12_rl_cal": {"easy": 66.7, "hard": 32.4, "original": 38.9}, |
| "v14": {"easy": 62.5, "hard": 30.6, "original": 40.3}, |
| "v15_teacher_rendered": {"easy_4096": 66.7, "hard_4096": 32.4, "original_4096": 40.3, "easy_6144": 66.7, "hard_6144": 32.4, "original_6144": 40.3} |
| }, |
| "v14": { |
| "train_rows": 62179, |
| "train_max_len": 6144, |
| "train_max_state": 5120, |
| "train_steps": 2017, |
| "validation_overlap_found": 422, |
| "validation_overlap_families": {"adj_route": 369, "adj_inject": 29, "adj_judge": 24}, |
| "clean_adjud_dev": {"n": 3378, "v12_rl_cal_accuracy": 0.370337, "v12_rl_cal_ece": 0.200263, "v14_accuracy": 0.812315, "v14_ece": 0.047555}, |
| "clean_mixed_dev": {"n": 10378, "v12_rl_cal_accuracy": 0.66323, "v12_rl_cal_ece": 0.047261, "v14_accuracy": 0.804105, "v14_ece": 0.033296} |
| }, |
| "teacher_rendering": { |
| "teacher": "Qwen3-14B", |
| "prompts": 600, |
| "fact_validated_rows": 478, |
| "rendered_state_file": "data/adjud_rendered.jsonl" |
| }, |
| "v15": { |
| "train_rows": 50324, |
| "dev_rows": 10417, |
| "train_dev_exact_overlap": 0, |
| "exact_public_jevbench_state_overlap": 0, |
| "train_steps": 1628, |
| "train_max_len": 6144, |
| "train_max_state": 5120, |
| "clean_dev_accuracy": 0.784967, |
| "clean_dev_ece": 0.026563, |
| "checkpoint_path": "runs/v15/final.pt", |
| "checkpoint_recovery": "Final save ran out of root-disk space; verified complete step1628.pt was renamed to final.pt. It loads successfully." |
| }, |
| "conclusions": [ |
| "Long-context adjudication training sharply improves leak-free procedural adjudication accuracy, but the gains transfer weakly to the authored JevBench hard tier.", |
| "The teacher-rendered mixture restores the base checkpoint's easy/hard public accuracy but does not improve the hard frontier; increasing eval context from 4096 to 6144 had no measurable public-tier effect.", |
| "The requested 75% easy / 60% hard target was not achieved." |
| ], |
| "artifacts": [ |
| "runs/v14/final.pt", |
| "runs/v14/leakfree_eval.json", |
| "runs/v14/jevbench_public_summary.json", |
| "runs/v14/jevbench_soup_public.json", |
| "runs/v15/final.pt", |
| "runs/v15/train.log", |
| "runs/v15/eval_dev.log", |
| "runs/v15/jevbench_public_4k.json", |
| "runs/v15/jevbench_public_6k.json", |
| "scripts/build_v15_mix.py", |
| "scripts/eval_jevbench_public.py", |
| "scripts/chain_v15.sh" |
| ] |
| } |
|
|
