---
license: apache-2.0
task_categories:
- video-text-to-text
language:
- en
---
# VideoChat3-Stage3-Training-Data

This repository includes all annotation files used across the four training stages of VideoChat3, from Stage 0 to Stage 3.

You can refer to the provided [source-data links](https://huggingface.co/datasets/MCG-NJU/VideoChat3-Training-Data-Annotations/tree/main/training_data_source_tables) to download videos, images, and other multimedia data for training. In `videochat3_data_annotations`, we also provide a `source` field to indicate the source dataset for each entry.

To facilitate Stage 3 training reproduction using the high-quality open-source datasets we collected and curated, we additionally provide a more detailed introduction at [VideoChat3-Stage3-Training-Data](https://huggingface.co/datasets/lmwang/VideoChat3-Stage3-Training-Data).

[📄 Paper](http://arxiv.org/abs/2607.14935) · [🌐 Homepage](https://mcg-nju.github.io/VideoChat3/) · [💻 GitHub](https://github.com/MCG-NJU/VideoChat3) · [🤗 Paper Page](https://huggingface.co/papers/2607.14935)

<p align="center">
  <img src="training_pipeline.jpg" alt="VideoChat3 Training Overview" width="100%">
</p>

## Data Organization

| Component | Description |
| --- | --- |
| `training_data_annotations` | Training annotation data for each stage |
| `videochat3_data_annotations` | Annotation files for individual datasets |
| `training_data_source_tables` | Source datasets and paths |


In `videochat3_data_annotations`, the Stage 0 and Stage 2 annotation data partially overlap with our previously released [VideoChat3-Academic2M](https://huggingface.co/datasets/MCG-NJU/VideoChat3-Academic2M) and [VideoChat3-LV116K](https://huggingface.co/datasets/MCG-NJU/VideoChat3-LV116K). For convenience of downloading and use, we re-upload them here.

> [!Note]
> Because the `training_data_source_tables` contain data from a wide range of sources, there may be occasional errors. If you spot any issues, please feel free to contact me at zzhiqiu997@gmail.com.

## Citation

If you use this data, please cite VideoChat3 and the original video datasets used by the annotations.

```
@misc{videochat3,
      title={VideoChat3: Fully Open Video MLLM for Efficient and Generalist Video Understanding}, 
      author={Xinhao Li and Yuhan Zhu and Xiangyu Zeng and Yuhao Dong and Haoning Wu and Zhiqiu Zhang and Yuandong Yang and Changlian Ma and Qingyu Zhang and Yansong Shi and Xinyu Chen and Haoran Chen and Zizheng Huang and Jun Zhang and Kun Ouyang and Lin Sui and Ziang Yan and Yicheng Xu and Chenting Wang and Yinan He and Hongjie Zhang and Yi Wang and Yu Qiao and Yali Wang and Ziwei Liu and Kai Chen and Limin Wang},
      year={2026},
      eprint={2607.14935},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.14935}, 
}
```