enmingzhangzz/Qwen2.5-VL-7B-OPSD-VisionZip-r010-B-Mean1Reweight-ForwardKL-gap01-original10240

Qwen2.5-VL-7B-OPSD-VisionZip-r010-B-Mean1Reweight-ForwardKL-gap01-original10240

Final PEFT/LoRA adapter from an OPSD experiment.

Variant

• —Objective: OPSD + full-token mean-normalized budget-JSD reweighting
• —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 1 x accumulation 8)
• —LoRA: r=16, alpha=32
• —Image pixels: 846720

• —Budget signal: B_t = JSD(P_student,b+gap || P_student,b)
• —B/B+ intervention gap: 0.01 (absolute vision-token retention ratio)
• —Token coverage: all valid rollout tokens
• —Detached token weight: w_t = B_t / mean_valid(B)
• —Per-response token-weight mean: 1
• —Training divergence: forward KL
• —Training objective: mean_valid[w_t * KL_t]

Files

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

Adapter SHA256: b4bed8526a05bcf047d1f154e892286300825345f0bb91178ef72b0227ad3061

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.