enmingzhangzz/Qwen2.5-VL-7B-OPSD-VisionZip-r010-balanced-10240

Qwen2.5-VL-7B-OPSD-VisionZip-r010-balanced-10240

Final PEFT/LoRA adapter from the balanced correct/incorrect OPSD experiment.

Variant

• —Objective: Standard OPSD
• —Base model: Qwen/Qwen2.5-VL-7B-Instruct
• —Training samples: 10240 (5,120 base-model-correct + 5,120 base-model-incorrect)
• —Pruning: official VisionZip
• —Vision-token retention ratio: 0.10
• —EMA teacher decay: 0.9999
• —Global batch size: 32 (4 GPUs x micro-batch 8)
• —LoRA: r=16, alpha=32
• —Image pixels: 846720

Files

adapter_model.safetensors and adapter_config.json are the final adapter at step 10240. Audit and reproducibility metadata are under training/.

Adapter SHA256: d972276d84eae934c545dbada8e5f25996e052e6a4390dcb3a38c87e2b4b2a2f

Load this adapter on top of Qwen/Qwen2.5-VL-7B-Instruct with PEFT. The VisionZip runtime patch used by the OPSD repository is still required for pruned inference.