ckpt-sudoku

Evolution-Strategies fine-tuning checkpoints from IBM Blue Vela.

iter<N>.pth are the periodic training checkpoints and final/pytorch_model.pth is the end-of-run save. The two are NOT copies of each other even at the same N -- the final save is written a few ES steps after the last periodic one, measured at a relative L2 difference of 0.002 to 0.004 where ten iterations move 0.008. Treat final/ as the canonical end-of-run weights.

One copy per iteration is published. Where a run has both a final save and a periodic checkpoint at that same iteration, only the final save is here. Periodic checkpoints at earlier iterations are kept, and a run that never reached its end keeps every periodic checkpoint it has.

All .pth are bf16 state dicts keyed by vLLM parameter names (fused qkv_proj, gate_up_proj), loadable by the trainer resume path.

run	base model	final iteration	periodic checkpoints
sudoku-k10-9b-fixed	Qwen/Qwen3.5-9B	500	none
sudoku-k10-9b-fresh	Qwen/Qwen3.5-9B	500	none
sudoku-k10-9b-hetero	Qwen/Qwen3.5-9B	500	none
sudoku-k10-9b-mirror	Qwen/Qwen3.5-9B	500	none
sudoku-k10-9b-mirror-v3	Qwen/Qwen3.5-9B	500	none
sudoku-k10-fixed	Qwen/Qwen3.5-4B	500	none
sudoku-k10-fresh	Qwen/Qwen3.5-4B	500	none
sudoku-k10-mirror	Qwen/Qwen3.5-4B	500	none
sudoku-k10-mirror-v3	Qwen/Qwen3.5-4B	500	none
sudoku-k15-9b-fixed	Qwen/Qwen3.5-9B	500	none
sudoku-k15-9b-fresh	Qwen/Qwen3.5-9B	500	none
sudoku-k15-9b-hetero	Qwen/Qwen3.5-9B	500	none
sudoku-k15-9b-mirror	Qwen/Qwen3.5-9B	500	none
sudoku-k15-9b-mirror-v3	Qwen/Qwen3.5-9B	500	none
sudoku-k15-fixed	Qwen/Qwen3.5-4B	500	none
sudoku-k15-fresh	Qwen/Qwen3.5-4B	500	none
sudoku-k15-hetero	Qwen/Qwen3.5-4B	500	none
sudoku-k15-mirror	Qwen/Qwen3.5-4B	500	none
sudoku-k15-mirror-v3	Qwen/Qwen3.5-4B	500	none