CapRL-Video-QA-20K.jsonl Video Path Setup

Each value is a relative path under the Hugging Face dataset root of lmms-lab/LLaVA-Video-178K.

Example:

"videos": ["0_30_s_youtube_v0_1/videos/liwei_youtube_videos/videos/youtube_video_2024/ytb_khSwLQOthHQ.mp4"]

Required Video Data

Download the original videos from Hugging Face:

• Dataset: https://huggingface.co/datasets/lmms-lab/LLaVA-Video-178K
• Required subdirectories for this 20k subset:
  • 0_30_s_youtube_v0_1
  • 0_30_s_academic_v0_1

On the Hugging Face file tree, these videos are distributed as compressed archives:

• 0_30_s_youtube_v0_1/0_30_s_youtube_v0_1_videos_*.tar.gz
• 0_30_s_academic_v0_1/0_30_s_academic_v0_1_videos_*.tar.gz

The annotation JSON files in those HF folders are not required by CapRL-Video-QA-20K.jsonl, but downloading the full folders is also fine.

Expected Directory Layout

After downloading and extracting the archives, organize files as follows:

/path/to/LLaVA-Video-178K/
├── 0_30_s_youtube_v0_1/
│   ├── 0_30_s_youtube_v0_1_videos_1.tar.gz
│   ├── ...
│   └── videos/
│       └── liwei_youtube_videos/
│           └── videos/
│               └── youtube_video_2024/
│                   └── ytb_*.mp4
└── 0_30_s_academic_v0_1/
    ├── 0_30_s_academic_v0_1_videos_1.tar.gz
    ├── ...
    └── videos/
        └── academic_source/
            ├── Charades/
            ├── NextQA/
            ├── activitynet/
            └── ...

The paths stored in CapRL-Video-QA-20K.jsonl should be joined with /path/to/LLaVA-Video-178K. For example:

from pathlib import Path
video_root = Path('/path/to/LLaVA-Video-178K')
relative_path = sample['videos'][0]
video_path = video_root / relative_path

Download and Extract Example

huggingface-cli download lmms-lab/LLaVA-Video-178K \
  --repo-type dataset \
  --local-dir /path/to/LLaVA-Video-178K \
  --include '0_30_s_youtube_v0_1/*' '0_30_s_academic_v0_1/*'
cd /path/to/LLaVA-Video-178K
for f in 0_30_s_youtube_v0_1/*_videos_*.tar.gz; do tar -xzf "$f" -C 0_30_s_youtube_v0_1; done
for f in 0_30_s_academic_v0_1/*_videos_*.tar.gz; do tar -xzf "$f" -C 0_30_s_academic_v0_1; done

If your downloader places files in a different location, keep the extracted videos/ directories under the two split folders shown above, or update your training script to join the relative paths with the actual dataset root.