Reward-Augmented Data Enhances Direct Preference Alignment of LLMs

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Zhang, Shenao, Liu, Zhihan, Liu, Boyi, Zhang, Yufeng, Yang, Yingxiang, Liu, Yongfei, Chen, Liyu, Sun, Tao, Wang, Zhaoran
Format: Preprint
Publié: 2024
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866909606726336512
author Zhang, Shenao
Liu, Zhihan
Liu, Boyi
Zhang, Yufeng
Yang, Yingxiang
Liu, Yongfei
Chen, Liyu
Sun, Tao
Wang, Zhaoran
author_facet Zhang, Shenao
Liu, Zhihan
Liu, Boyi
Zhang, Yufeng
Yang, Yingxiang
Liu, Yongfei
Chen, Liyu
Sun, Tao
Wang, Zhaoran
contents Preference alignment in Large Language Models (LLMs) has significantly improved their ability to adhere to human instructions and intentions. However, existing direct alignment algorithms primarily focus on relative preferences and often overlook the qualitative aspects of responses, despite having access to preference data that includes reward scores from judge models during AI feedback. Striving to maximize the implicit reward gap between the chosen and the slightly inferior rejected responses can cause overfitting and unnecessary unlearning of the high-quality rejected responses. The unawareness of the reward scores also drives the LLM to indiscriminately favor the low-quality chosen responses and fail to generalize to optimal responses that are sparse in data. To overcome these shortcomings, our study introduces reward-conditioned LLM policies that discern and learn from the entire spectrum of response quality within the dataset, helping extrapolate to more optimal regions. We propose an effective yet simple data relabeling method that conditions the preference pairs on quality scores to construct a reward-augmented dataset. The experiments across various benchmarks and diverse models demonstrate that our approach consistently boosts DPO by a considerable margin. Through comprehensive ablation studies, we demonstrate that our method not only maximizes the utility of preference data but also mitigates the issue of unlearning, demonstrating its broad effectiveness beyond mere data expansion. Our code is available at https://github.com/shenao-zhang/reward-augmented-preference.
format Preprint
id arxiv_https___arxiv_org_abs_2410_08067
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Reward-Augmented Data Enhances Direct Preference Alignment of LLMs
Zhang, Shenao
Liu, Zhihan
Liu, Boyi
Zhang, Yufeng
Yang, Yingxiang
Liu, Yongfei
Chen, Liyu
Sun, Tao
Wang, Zhaoran
Machine Learning
Artificial Intelligence
Preference alignment in Large Language Models (LLMs) has significantly improved their ability to adhere to human instructions and intentions. However, existing direct alignment algorithms primarily focus on relative preferences and often overlook the qualitative aspects of responses, despite having access to preference data that includes reward scores from judge models during AI feedback. Striving to maximize the implicit reward gap between the chosen and the slightly inferior rejected responses can cause overfitting and unnecessary unlearning of the high-quality rejected responses. The unawareness of the reward scores also drives the LLM to indiscriminately favor the low-quality chosen responses and fail to generalize to optimal responses that are sparse in data. To overcome these shortcomings, our study introduces reward-conditioned LLM policies that discern and learn from the entire spectrum of response quality within the dataset, helping extrapolate to more optimal regions. We propose an effective yet simple data relabeling method that conditions the preference pairs on quality scores to construct a reward-augmented dataset. The experiments across various benchmarks and diverse models demonstrate that our approach consistently boosts DPO by a considerable margin. Through comprehensive ablation studies, we demonstrate that our method not only maximizes the utility of preference data but also mitigates the issue of unlearning, demonstrating its broad effectiveness beyond mere data expansion. Our code is available at https://github.com/shenao-zhang/reward-augmented-preference.
title Reward-Augmented Data Enhances Direct Preference Alignment of LLMs
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2410.08067