The dataset is split into sharded .tar.gz archives (~1000 video+caption pairs each).
Each shard has the following structure:
shard_XXXX.tar.gz
└── shard_XXXX/
├── videos/
│ ├── <video_id>_clip.mp4
│ └── ...
└── captions/
├── <video_id>_clip.txt
└── ...
videos/: 5-second 1080p MP4 clips with audio
captions/: Corresponding text caption for each video clip
