ckpt-math

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.

The two math-l5 runs v1 and v3 carry their periodic checkpoints on a 50-iteration grid. Their prune ran at the START of each dispatch, so the final dispatch of each was never pruned and left a dense every-ten tail of different length in each. Keeping multiples of 50 applies the same policy the runs applied to themselves earlier and makes the two comparable.

Runs still training are not here yet. They are published once they reach their target and write a final save, so that no iteration ever appears twice.

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
math7b-fixed	Qwen/Qwen2.5-Math-7B	500	none
math7b-fresh	Qwen/Qwen2.5-Math-7B	500	none
math7b-hetero	Qwen/Qwen2.5-Math-7B	500	none
math7b-mirror	Qwen/Qwen2.5-Math-7B	500	none
math7b-mirror-v3	Qwen/Qwen2.5-Math-7B	500	none
math-l5-v1-m76-seed42	Qwen/Qwen3-4B-Instruct-2507	500	100, 150, 200, 250, 300, 350, 400, 450, 50
math-l5-v3-m76-seed42	Qwen/Qwen3-4B-Instruct-2507	500	100, 150, 200, 250, 300, 350, 400, 450, 50
or1-hetero	Qwen/Qwen3-4B-Instruct-2507	500	none
or1-fresh	Qwen/Qwen3-4B-Instruct-2507	500	none