Shanghai Artificial Intelligence Laboratory

[Paper] [Data(huggingface)] [Data(opendatalab)] [Models]

We propose HA-DPO (Hallucination-aware Direct Preference Optimization), which eliminates LVLM hallucination using GPT-4 generated positive-negative data based on DPO (Direct Preference Optimization).

This repo provides:

• High-quality positive-negative hallucination-aware data

• DPO framework for multiple LVLM (based on TRL)

  • MiniGPT-4
  • LLaVA-1.5
  • InstructBLIP

• SHR Evaluation (Sentence-level Hallucination Ratio)

• Human-free positive-negative data mining pipeline

Online demo will be available soon!

1. create a new conda environment

conda create -n hadpo python==3.9

2. install requirements

conda activate hadpo
pip install -e .

For data preparation, please refer to data preparation.

For model training and evaluation, please refer to docs in following:

For SHR Evaluation, please refer to SHR Evaluation.

• MiniGPT-4. The MiniGPT-4 part of HA-DPO is based on the official MiniGPT-4 implementation.
• VIGC. The InstructBLIP part of HA-DPO is built on VIGC, which is an amazing visual instruction generation and correction method.
• LLaVA-1.5. The LLaVA-v1.5 part of HA-DPO is based on the official LLaVA-1.5 implementation, which is a great open-source work on LVLM.
• TRL. Most model training and optimizing codes of HA-DPO are stemed from TRL, which is a great human-preference learning framework on LLM.

You can find more details in our paper.

If you're using HA-DPO in your research or applications, please cite using this BibTeX:

@misc{zhao2023hallucinations,
      title={Beyond Hallucinations: Enhancing LVLMs through Hallucination-Aware Direct Preference Optimization}, 
      author={Zhiyuan Zhao and Bin Wang and Linke Ouyang and Xiaoyi Dong and Jiaqi Wang and Conghui He},
      year={2023},
      eprint={2311.16839},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

If you have any questions, comments or suggestions, please do not hesitate to contact us at [email protected] and [email protected].

Apache License 2.0