Dataset Setup
Download status for POPE + MME benchmark datasets. Follow the steps below if a dataset is marked Not ready.
Checking dataset status...
Model Weights
Qwen/Qwen2.5-VL-3B-Instruct (~7 GB) downloads automatically on first smoke-test run. Alternatively, pre-download with:
huggingface-cli download Qwen/Qwen2.5-VL-3B-Instruct
If RTX 3050 OOM at float16 (~5.5 GB), set load_in_4bit: true in visual_perception/config/default_config.yaml (~2 GB, requires bitsandbytes).