[AAAI 2025] Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback

Wenyi Xiao^1* , Ziwei Huang^1* , Leilei Gan^1† , Wanggui He²
Haoyuan Li² , Zhelun Yu² , Fangxun Shu² , Hao Jiang² , Linchao Zhu^1
1 Zhejiang University      ² Alibaba Group      
^*Equal contribution        ^†Corresponding author

Overview

This repository contains the official implementation of the paper "Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback".

Getting Started

Setup

git clone https://github.com/Mr-Loevan/HSA-DPO.git
cd HSA-DPO
pip install -r requirements.txt

Dataset

pip install -U huggingface_hub
huggingface-cli download --repo-type dataset WenyiXiao/HSA-DPO

For hallucination detection: The image is sourced from Visual Genome, and the training dataset can be found in hsa_dpo_detection.jsonl.

For hallucination mitigation: The image is located in hsa_dpo_imgs.tar.gz, and the preferences dataset is available in hsa_dpo_preference_llava1dot5.jsonl. Note that in llava1dot5, 'rejected' is generated by llava-v1.5.

Model LoRA Weight

pip install -U modelscope
modelscope download --model xiaowenyi/HSA-DPO

Refer to Instructions to install inference requirements and use inference code.

Training Code

The code is currently undergoing internal review. Please stay tuned!

Todo List

• paper
• detection & mitigation datasets
• model weights
• training code