Base: Qwen/Qwen2.5-VL-7B-Instruct
Method: TTRL (self-label majority-vote), MM-UPT-aligned recipe (KL 1e-2/temp 0.7/gen 10/len 2048), 1 epoch
Dataset: mmr1 (~8k), 1 epoch
Checkpoint: best (best_model)
4-bench avg: **49.69 **
best repo: q1716523669/mllm-mmr1-ttrl-qwen25vl7b-mmupt · endpoint repo: q1716523669/mllm-mmr1-ttrl-qwen25vl7b-mmupt-endpoint
eval: prompt=answer, greedy T=0, mathruler; see results.csv in run.