prepare_lvbench.md

Prepare LVBench Dataset

Step 1: download LVBench data from huggingface

git clone https://huggingface.co/datasets/THUDM/LVBench # Contain annotations only
git clone https://huggingface.co/datasets/AIWinter/LVBench # Contain videos only

Move all_files in AIWinter/LVBench into THUDM/LVBench.

Denote the root directory of download LVBench dataset as lvbench_root, it should has the following structure:

${lvbench_root}/
├── docs/
├── video_info.meta.jsonl
├── all_videos_split.zip.001
├── all_videos_split.zip.002
├── ...
└── all_videos_split.zip.014

Step 2: Unzip everything

cd ${lvbench_root}
cat all_videos_split.zip.* > all_videos.zip
unzip all_videos.zip

Step 3: Build LVBench dataset

cd evaluation
python scripts/utils/build_lvbench_dataset.py --hf_root ${lvbench_root}

The file structure is as follows:

evaluation/
├── dataset/
    ├── lvbench/
        ├── lvbench.json
├── ...