LLaVA-Hound-DPO

<br> <br>

LLaVA-Hound Model Card

Model details

Model type: LLaVA-Hound is an open-source video large multimodal model, fine-tuned from video instruction following data based on large language model.

This model is trained with DPO from 17k video instruction preference dataset from ShareGPTVideo/LLaVA-Hound-SFT ckpt.

Base LLM: lmsys/vicuna-7b-v1.5

Model date: Trained on March 15, 2024.

Paper or resources for more information: https://github.com/RifleZhang/LLaVA-Hound-DPO

License

lmsys/vicuna-7b-v1.5 license.

Where to send questions or comments about the model: https://github.com/RifleZhang/LLaVA-Hound-DPO/issues

Intended use

Primary intended uses: Video (image) instruction-following.

Primary intended users: Researchers in artificial intelligence, large multimodal model, etc.

Training dataset

ShareGPTVideo dataset.

Evaluation

Follow https://github.com/RifleZhang/LLaVA-Hound-DPO/blob/main/README.md

Paper

https://huggingface.co/papers/2404.01258

citation

@article{zhang2024direct,
  title={Direct Preference Optimization of Video Large Multimodal Models from Language Model Reward},
  author={Zhang, Ruohong and Gui, Liangke and Sun, Zhiqing and Feng, Yihao and Xu, Keyang and Zhang, Yuanhan and Fu, Di and Li, Chunyuan and Hauptmann, Alexander and Bisk, Yonatan and others},
  journal={arXiv preprint arXiv:2404.01258},
  year={2024}
}