Uniform index for the M5 training ablation: Qwen3.5-0.8B trained with GRPO + a Wikipedia search tool on MuSiQue, varying one axis at a time, reward shape (F1-only / F1+format / EM-only) across three seeds (42 / 43 / 44). The full checkpoints (every 10 steps) + per-step rollouts live in the 9 existing per-run repos below (this index does NOT duplicate the ~1.9 TB of weights; it documents them uniformly… See the full description on the dataset page:
https://huggingface.co/datasets/pantomiman/qwen3.5-0.8b-grpo-musique-m5.
