FlowNar-Data

Dataset features and annotations for FlowNar: Scalable Streaming Narration for Long-Form Videos.

FlowNar-Data provides pre-extracted video features and narration annotations used to train and evaluate FlowNar on long-form streaming video narration. It covers three benchmarks: Ego4D, EgoExo4D, and EpicKitchens100.

Dataset Structure

FlowNar-Data/
├── ek100/
│   ├── annotations/
│   ├── features/
│   └── features_metadata.json
├── ego4d/
│   ├── annotations/
│   ├── features/
│   └── features_metadata.json
└── egoexo/
    ├── annotations/
    ├── features/
    └── features_metadata.json

The feature data is distributed as multi-part tar archives. Reassemble and extract each archive before use:

cat <archive_name>.tar.* | tar -xf -

Usage

Download the dataset from Hugging Face:

huggingface-cli download zeyun-zhong/FlowNar-Data \
    --repo-type dataset \
    --local-dir /path/to/FlowNar-Data

Use the downloaded path as --data_root when running FlowNar training or streaming generation scripts.

Intended Use

This dataset is intended for research on long-form video understanding, streaming video narration, video-text-to-text modeling, and self-conditioned evaluation of video-language models.

Related Resources

• Paper: https://arxiv.org/abs/2606.00620
• Project page: https://zeyun-zhong.github.io/FlowNar/
• Code: https://github.com/zeyun-zhong/flownar

License

The FlowNar code is released under the MIT License. This dataset contains processed features and annotations derived from Ego4D, EgoExo4D, and EpicKitchens100; users should also comply with the terms of the original datasets.

Citation

@inproceedings{zhong2026flownar,
  title     = {{FlowNar}: Scalable Streaming Narration for Long-Form Videos},
  author    = {Zhong, Zeyun and Martin, Manuel and Wu, Chengzhi and Schneider, David and
               Diederichs, Frederik and Gall, Juergen and Beyerer, Juergen},
  booktitle = {International Conference on Machine Learning},
  year      = {2026},
  publisher = {PMLR},
  note      = {Accepted, to appear}
}