enmingzhangzz/Qwen2.5-VL-7B-SFT-VisionZip-r010-original10240

Qwen2.5-VL-7B-SFT-VisionZip-r010-original10240

Final PEFT/LoRA adapter from supervised fine-tuning of Qwen2.5-VL-7B with official VisionZip visual-token pruning.

Training configuration

• —Base model: Qwen/Qwen2.5-VL-7B-Instruct (fresh initialization; no OPSD/EPIC adapter)
• —Objective: assistant-response supervised cross-entropy
• —Training samples: 10,240 from the fixed original OpenMMReasoner prefix (not outcome-balanced)
• —Pruner: official VLMEvalKit VisionZip implementation
• —Fixed visual-token retention ratio: 0.10
• —Smoke-measured visual-token retention ratio: 0.09934522
• —Parameter scope: language-decoder-only LoRA
• —LoRA: r=16, alpha=32, dropout=0.0
• —Global batch: 32 (4 GPUs x micro-batch 8 x accumulation 1)
• —Learning rate: 2e-05
• —Image pixels: 846720 to 846720
• —Prompt mode: reasoning (<analysis>...</analysis><answer>...</answer>)

The resolved configuration, complete rank-0 training log, smoke audit, validation report, and trainable-parameter report are included under training/.

Adapter SHA256: 9463250e17a196055c2f54e205587c7e727c3b5b6b7a30eb3158c5967ccb9f8e

Load this adapter on top of Qwen/Qwen2.5-VL-7B-Instruct with PEFT. The same VisionZip runtime patch is required for 10% pruned inference.