CoRL · Qwen2.5-VL-7B-Instruct · GT-GRPO (ground-truth verifiable reward) · mmr1 (best)

• Base: Qwen/Qwen2.5-VL-7B-Instruct
• Method: GT-GRPO (ground-truth verifiable reward)
• Dataset: mmr1 (~8k), 1 epoch = 722 steps
• This checkpoint: BEST (best-by-val, step 580) — selected by in-loop MathVista-150 eval_reward
• Paper cell: MLLM main table (mmr1, big tier) · Qwen2.5-VL-7B-Instruct · GT
• WandB run: https://wandb.ai/logan-yang2002-johns-hopkins-university/co-rl-mllm-mmr1/runs/qwen25vl7b_gt_mmr1-v1
• best repo: q1716523669/mllm-mmr1-gt-qwen25vl-7b · endpoint repo: q1716523669/mllm-mmr1-gt-qwen25vl-7b-endpoint

Eval — 4 benchmarks (prompt=answer · greedy T=0 · mathruler)