COF readiness-v3 artifacts (sliding / sudoku / tangram)
Checkpoints and train LMDBs from the 2026-08-30..09-03 debug of the three
GPT-built visual-reasoning tasks in mianwu01/COF-maze
(branch review/cof-6k-snapshot-20260825, commit fbee115).
Code, configs, exact data-generation commands, evaluation summaries, and the
full analysis live in that branch:
reports/2026-08-30_reasoning_readiness_debug.md— why the readiness gates failed and what fixes itreports/2026-09-03_readiness_v3_cluster_handoff.md— how to restore and continue on a clustercausal-forcing-changes/modifications.patch— trainer overlay incl. the remediation-v3 loss controls
All checkpoints are Wan2.1-T2V-1.3B causal (AR, I2V) Causal-Forcing generators,
image_or_video_shape [1,21,16,60,104], trained with the v3 recipe from the
public maze init Aelien/COF-maze-checkpoints/maze_ar_cont/checkpoint_model_001500.
model.pt holds generator and generator_ema; trainer.pt (where present)
adds optimizer/data/RNG state for exact resume.
| path | what | dev8 result (seed 101) |
|---|---|---|
checkpoints/sudoku_maze_init_v3_long2/global_1985/{model,trainer}.pt |
sudoku 9x9 hidden-endpoint, exact-resume point | 1/8 exact, 50% valid progress, givens 8/8 |
checkpoints/sudoku_maze_init_v3_long2/global_1697/model.pt |
sudoku, best valid-progress model | 1/8 exact, 63% valid progress, 50% legal |
checkpoints/sliding_maze_init_v3_x4/step_001200/{model,trainer}.pt |
sliding 3x3, 1024-sample run, resume point | evals pending |
checkpoints/sliding_maze_init_v3_long/step_000600/model.pt |
sliding 3x3, 256-sample measured peak | 81% legal, 83% valid state, 4.6 moves |
data/reasoning_tasks/<task>/lmdb/train (+ sliding_exact_v1/lmdb/train1024) |
fp16 VAE latents (256 rows per task; 1024-row sliding set), action masks baked in | -- |
Configs: experiments/configs/{sudoku_maze_init_v3_long2,sliding_maze_init_v3_x4,sliding_maze_init_v3_long}.yaml.
Evaluate with scripts/readiness_v3/eval_chain.sh; continue training with scripts/readiness_v3/run_segments.sh.