LongVT-Source

This repository contains the source video and image files for the LongVT project.

Overview

LongVT is an end-to-end agentic framework that enables "Thinking with Long Videos" via interleaved Multimodal Chain-of-Tool-Thought. This dataset provides the raw media files referenced by the training annotations in LongVT-Parquet.

Dataset Structure

The source files are organized by dataset type and stored as zip archives:

Training Data

Evaluation Data

We have transferred the source videos of VideoSIAH-Eval to longvideotool/VideoSIAH-Eval.

Download

Install huggingface_hub

pip install huggingface_hub

Download all source files

huggingface-cli download longvideotool/LongVT-Source --repo-type dataset --local-dir ./source

Or download specific files

huggingface-cli download longvideotool/LongVT-Source longvideoreason_1.zip --repo-type dataset --local-dir ./source## Usage

After downloading, extract the zip files to obtain the source media:

cd source unzip "*.zip"The extracted paths will match those referenced in the LongVT-Parquet annotations.

Related Resources

• 📄 Paper: arXiv:2511.20785
• 🌐 Project Page: LongVT Website
• 💻 Code: GitHub Repository
• 📊 Annotations: LongVT-Parquet
• 🤗 Models: LongVT Collection

Citation

If you find LongVT useful for your research and applications, please cite using this BibTeX:

@misc{yang2025longvtincentivizingthinkinglong,
      title={LongVT: Incentivizing "Thinking with Long Videos" via Native Tool Calling}, 
      author={Zuhao Yang and Sudong Wang and Kaichen Zhang and Keming Wu and Sicong Leng and Yifan Zhang and Bo Li and Chengwei Qin and Shijian Lu and Xingxuan Li and Lidong Bing},
      year={2025},
      eprint={2511.20785},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.20785}, 
}

License

This dataset is released under the Apache 2.0 License.