yinyueqin / sapo Goto Github PK

This repository provides the official PyTorch implementation for the following paper:

Self-Augmented Preference Optimization: Off-Policy Paradigms for Language Model Alignment [Arxiv]
Yueqin Yin*, Zhendong Wang*, Yujia Xie, Weizhu Chen and Mingyuan Zhou
(* denotes equal contribution)
The University of Texas At Austin, Microsoft Azure AI

Abstract: Traditional language model alignment methods, such as Direct Preference Optimization (DPO), are limited by their dependence on static, pre-collected paired preference data, which hampers their adaptability and practical applicability. To overcome this limitation, we introduce Self-Augmented Preference Optimization (SAPO), an effective and scalable training paradigm that does not require existing paired data. Building on the self-play concept, which autonomously generates negative responses, we further incorporate an off-policy learning pipeline to enhance data exploration and exploitation. Specifically, we employ an Exponential Moving Average (EMA) model in conjunction with a replay buffer to enable dynamic updates of response segments, effectively integrating real-time feedback with insights from historical data. Our comprehensive evaluations of the LLaMA3-8B and Mistral-7B models across benchmarks, including the Open LLM Leaderboard, IFEval, AlpacaEval 2.0, and MT-Bench, demonstrate that SAPO matches or surpasses established offline contrastive baselines, such as DPO and Odds Ratio Preference Optimization, and outperforms offline self-play methods like SPIN.

1. Clone this repo:

   git clone https://github.com/yinyueqin/SAPO.git
   cd SAPO

2. Install dependent packages: A suitable Anaconda environment named sapo can be created and activated with:

   conda create -n sapo python=3.10
   conda activate sapo
   python -m pip install .
   python -m pip install flash-attn --no-build-isolation
   

For detailed script commands, please refer to the scripts in scripts/sapo_dpo.sh and scripts/sapo_orpo.sh.

If you find this work useful for your research, please consider citing our paper:

@article{yin2024self,
  title={Self-Augmented Preference Optimization: Off-Policy Paradigms for Language Model Alignment},
  author={Yin, Yueqin and Wang, Zhendong and Xie, Yujia and Chen, Weizhu and Zhou, Mingyuan},
  journal={arXiv preprint arXiv:2405.20830},
  year={2024}
}

This repo is built upon SPIN and TRL. We thank the authors for their excellent work.