| { |
| "session": "v12_adversarial_round", |
| "date": "2025-09-24", |
| "pod": "gpu-518b1253 A40 48GB", |
| "base_checkpoint": "v11_anneal", |
| "strategy": "adversarial/compositional reasoning data instead of templates: BIG-Bench Hard-style tasks + ReClor + brute-force-verified puzzle generator (gen_hard2)", |
| "data": { |
| "v12_mix_rows": 231386, |
| "v12_dev_rows": 10107, |
| "new_sources": [ |
| "tasksource/bigbench 16 configs", |
| "reclor 4.6k", |
| "gh2 22.5k solver-verified puzzles" |
| ], |
| "anneal_mix_rows": 69893 |
| }, |
| "results": { |
| "v11_anneal_baseline_v12dev": { |
| "acc": 0.614, |
| "ece": 0.079, |
| "noul_brier": 0.133 |
| }, |
| "v12": { |
| "dev_acc": 0.7336, |
| "dev_ece": 0.039, |
| "noul_brier": 0.034, |
| "tev1_transfer": 0.438, |
| "gh2_dev": 0.504, |
| "reason_dev": 0.648, |
| "jevbench": { |
| "easy": 0.583, |
| "hard": 0.306, |
| "original": 0.444 |
| } |
| }, |
| "v12_anneal": { |
| "dev_acc": 0.7329, |
| "dev_ece": 0.034, |
| "noul_brier": 0.033, |
| "jevbench": { |
| "easy": 0.604, |
| "hard": 0.324, |
| "original": 0.458 |
| } |
| }, |
| "v12_soup_50v12_30a_20v10": { |
| "jevbench": { |
| "easy": 0.604, |
| "hard": 0.342, |
| "original": 0.431 |
| } |
| }, |
| "v12a_soup50": { |
| "jevbench": { |
| "easy": 0.667, |
| "hard": 0.315, |
| "original": 0.431 |
| } |
| }, |
| "v12a_soup40": { |
| "jevbench": { |
| "easy": 0.667, |
| "hard": 0.342, |
| "original": 0.403 |
| }, |
| "dev_acc": 0.6487, |
| "dev_ece": 0.021 |
| }, |
| "v12_rl_grpo400": { |
| "data": "12k focused exact-match families (order/swaps/tracking/logical/navigate)", |
| "reward_mean": 0.42, |
| "gh2_dev": 0.502, |
| "reason_dev": 0.639, |
| "jevbench": { |
| "easy": 0.604, |
| "hard": 0.351, |
| "original": 0.486 |
| }, |
| "note": "BEST original tier of all checkpoints (48.6), hard 35.1 near v10; calibration degraded (dev ECE 0.247)" |
| }, |
| "v12_rlsoup_35rl_35a_30v10": { |
| "jevbench": { |
| "easy": 0.646, |
| "hard": 0.342, |
| "original": 0.444 |
| } |
| } |
| }, |
| "heldout_family_gains_v12": { |
| "date_understanding": [ |
| 0.167, |
| 1.0 |
| ], |
| "temporal_sequences": [ |
| 0.625, |
| 1.0 |
| ], |
| "disambiguation_qa": [ |
| 0.778, |
| 1.0 |
| ], |
| "gh2_arith": [ |
| 0.116, |
| 0.905 |
| ], |
| "reclor": [ |
| 0.753, |
| 0.848 |
| ], |
| "hyperbaton": [ |
| 0.546, |
| 0.816 |
| ], |
| "gh2_comp_policy_brier": [ |
| 0.149, |
| 0.001 |
| ], |
| "gh2_temp_overlap_brier": [ |
| 0.705, |
| 0.095 |
| ] |
| }, |
| "unchanged_weak_families": [ |
| "tracking_shuffled_objects 0.16", |
| "gh2_swaps 0.22", |
| "gh2_order 0.28", |
| "logical_deduction 0.48" |
| ], |
| "conclusion": "v12/v12_anneal are the strongest general decision models (73.3% dev, best calibration, best JevBench original 45.8). nextrun_v10 remains best on JevBench easy (68.8). v12a_soup40 blends to 66.7 easy/34.2 hard while keeping most capability gains. Multi-step state tracking remains unsolved. GRPO on verifiable families lifted JevBench original to a session-best 48.6 and hard to 35.1 but hurt calibration; soups partially restore it. Release candidates: nextrun_v10 (bench easy), v12a_soup40 (balanced), v12_anneal (general+calibration), v12_rl (original/hard tier).", |
| "checkpoints": { |
| "runs/v12/final.pt": "full adversarial SFT from v11_anneal, 231k rows", |
| "runs/v12_anneal/final.pt": "replay-heavy low-LR anneal of v12, 70k rows", |
| "runs/v12a_soup40.pt": "0.4*v12_anneal + 0.6*nextrun_v10 weight soup", |
| "runs/v12a_soup50.pt": "0.5/0.5 soup", |
| "runs/v12_soup.pt": "0.5*v12+0.3*v11_anneal+0.2*v10", |
| "runs/v12_rl/step400.pt": "GRPO 400 steps on exact-match reasoning families from v12_anneal; best original tier", |
| "runs/v12_rlsoup.pt": "0.35*v12_rl+0.35*v12_anneal+0.30*v10" |
| } |
| } |
