Qwen2.5-VL-7B-Instruct trained for 30 Vision-OPD optimizer steps. The student retains a replayable random 5% subset of post-merge visual tokens, while the EMA teacher uses the full visual-token sequence.

Training uses Vision-OPD-6K with 96 prompts and 8 on-policy rollouts per step, a learning rate of 2e-6, and an EMA teacher update rate of 0.05.

Standard Transformers/vLLM loads the full-token model. RandomDrop5 inference requires the pruning-aware serving code in prune-opd.

Training data: Vision-OPD-6K.

zhuqiang/Vision-OPD-Qwen2.5-VL-7B-Instruct-RandomDrop5-Step30

Vision-OPD Qwen2.5-VL-7B RandomDrop5 Step30

zhuqiang/
Vision-OPD-Qwen2.5-VL-7B-Instruct-RandomDrop5-Step30