---
license: llama2
datasets:
- wangyueqian/HawkEye-IT
- wangyueqian/InternVid-G
- OpenGVLab/VideoChat2-IT
language:
- en
pipeline_tag: visual-question-answering
---
# <div style="display: flex; align-items: center;"> <img src="https://github.com/yellow-binary-tree/HawkEye/blob/main/assets/hawk.png?raw=True" alt="logo" width="50" height="50" style="margin: 0 10;"> <span style="margin: 10 10;">&emsp;HawkEye: Training Video-Text LLMs for Grounding Text in Videos</span> </div>

This repo provides the checkpoint of HawkEye, and our implementation of VideoChat2.

`videochat2-stage3-our_impl.pth` is the chekepoint of our reproduce of VideoChat2. You can use it as an substitution of `hawkeye.pth`.
- The difference between it and HawkEye is: not trained with data from [InternVid-G](https://github.com/yellow-binary-tree/HawkEye/blob/main/internvid_g/README.md).
- The difference between it and the original implementation of VideoChat2 is: the visual encoder is frozen, and not trained with image data from [VideoChat2-IT](https://github.com/OpenGVLab/Ask-Anything/blob/main/video_chat2/DATA.md)

For more details please refer to our [paper](https://arxiv.org/abs/2403.10228) and [github](https://github.com/yellow-binary-tree/HawkEye).