---

News

• Our AMP is accepted by NeurIPS 2024 as poster presentation!
• [2024/05/29] We relase AMP in arxiv! Our code, MRHal Benchmark, and models are now open source!

Overview

We present an automated Multi-level Preference (AMP) framework for Reinforcement Learning from Human Feedback (RLHF), which generates the high-quality multi-level preference dataset without any human/AI annotators and employs multi-level DPO (MDPO) algorithm. Our AMP achieves SOTA performance across multiple hallucination benchmarks, including MMHal-Bench, MRHal-Bench, LLaVA-Bench, and POPE.

Prepare

1. Install some important packages.

conda create -n amp python=3.10 -y
conda activate amp
pip install --upgrade pip
pip install -r requirements.txt

2. Download Base Model

   llava-7b-base

   llava-13b-base

Train

1. Prepare data from [RLHF-V], [SILKIE], [ShareGPT4V].

2. Download Data from this link.

3. Run the following code

sh scripts/13b-v1.5/train_dpo.sh    # 13B
sh scripts/7b-v1.5/train_dpo.sh     # 7B

Evaluation

MMHal-Bench

1. Download data from [MMHal-Bench].
2. Run the script

sh eval/eval_scripts/eval_mmhal.sh

MRHal-Bench

1. Download data from [MRHal-Bench].
2. Run the script

sh eval/eval_scripts/eval_mrhal.sh

LLaVA-Bench

1. Download data from [LLaVA-Bench] and [COCO] images.
2. Run the script

sh eval/eval_scripts/eval_pope.sh

POPE

1. Download data from [POPE] and [COCO] images.
2. Run the script

sh eval/eval_scripts/eval_llavab.sh

Model Zoo

You can also use our trained models for evaluation. We provide the lora adpater of each version.

Size	Dataset	Link
7B	MEG	MEG-7B
7B	IG	IG-7B
13B	MEG	MEG-13B
13B	IG	IG-13B

Dialogue Example

We provide several dialogue examples, with additional results available in the paper.

Citation

If you find this repository is useful, please consider star🌟 this repo and cite🖇️ our paper.

@article{zhang2024amp,
      title={Automated Multi-level Preference for MLLMs}, 
      author={Zhang, Mengxi and Wu, Wenhao and Yu, Lu and Song, Yuxin and Rong, Kang and Yao, Huanjin and Zhang, Jianbo and Liu, Fanglong and Feng, Haocheng and Sun, Yifan and Wang, Jingdong},
      journal={Advances in Neural Information Processing Systems},
      year={2024}
}

Thanks

Our code is partly based on [LLaVA], [LLaVA-RLHF], and [TRL]. Thanks for their excllent work!