enmingzhangzz/Qwen2.5-VL-7B-OPSD-VisionZip-r010-Btop50-absFbottom50-lambda05-gap01-10240

Qwen2.5-VL-7B-OPSD-VisionZip-r010-Btop50-absFbottom50-lambda05-gap01-10240

Final PEFT/LoRA adapter from an OPSD experiment.

Variant

• —Objective: OPSD + top-B/bottom-|F| grouped weighting
• —Base model: Qwen/Qwen2.5-VL-7B-Instruct
• —Training samples: 10240
• —Dataset tag: openmmreasoner_llava_cot_exact_prefix10240_decontam_v1_seed42
• —Balanced base-outcome sampling: false
• —Pruning: visionzip
• —Vision-token retention ratios: [0.1]
• —EMA teacher decay: 0.9999
• —Global batch size: 32 (4 GPUs x micro-batch 2 x accumulation 4)
• —LoRA: r=16, alpha=32
• —Image pixels: 846720

• —Grouping mode: token_budget_jsd_top50_abs_f_bottom50_grouped
• —B/B+ intervention: 0.01 (absolute retention ratio)
• —Top-B candidate fraction: 0.5
• —Bottom-|F| fraction within top-B: 0.5
• —Selected-group aggregation lambda: 0.5

Files

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

Adapter SHA256: 49051a9fd6a58ca9a3e20cf0359f738a8b24887cbf7354924a7dd9f4a9ccc5f4

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.