8 independent GRPO runs (seeds 0-7) finetuning Qwen/Qwen3-0.6B on gsm8k, saved at training steps 1, 4, 8, 12, 200. Weights only (no optimizer state).

Each checkpoint is a standard HF model dir; load with:

These accompany a study of the lottery gap (union vs. average accuracy across seeds), checkpoint divergence in weight space, and model merging / souping. checkpoint-200 is the final model; checkpoint-1/4/8/12 trace the early trajectory.