Saved in:
Bibliographic Details
Main Authors: Shen, Yaojie, Wang, Xinyao, Niu, Yulei, Zhou, Ying, Tang, Lexin, Zhang, Libo, Chen, Fan, Wen, Longyin
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2409.08845
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866929499591933952
author Shen, Yaojie
Wang, Xinyao
Niu, Yulei
Zhou, Ying
Tang, Lexin
Zhang, Libo
Chen, Fan
Wen, Longyin
author_facet Shen, Yaojie
Wang, Xinyao
Niu, Yulei
Zhou, Ying
Tang, Lexin
Zhang, Libo
Chen, Fan
Wen, Longyin
contents Preference Optimization (PO), is gaining popularity as an alternative choice of Proximal Policy Optimization (PPO) for aligning Large Language Models (LLMs). Recent research on aligning LLMs iteratively with synthetic or partially synthetic data shows promising results in scaling up PO training for both academic settings and proprietary trained models such as Llama3. Despite its success, our study shows that the length exploitation issue present in PO is even more severe in Iterative Preference Optimization (IPO) due to the iterative nature of the process. In this work, we study iterative preference optimization with synthetic data. We share the findings and analysis along the way of building the iterative preference optimization pipeline. More specifically, we discuss the length exploitation issue during iterative preference optimization and propose our training objective for iterative preference optimization, namely Agreement-aware Iterative Preference Optimization (AIPO). To demonstrate the effectiveness of our method, we conduct comprehensive experiments and achieve state-of-the-art performance on MT-Bench, AlpacaEval 2.0, and Arena-Hard. Our implementation and model checkpoints will be made available at https://github.com/bytedance/AIPO.
format Preprint
id arxiv_https___arxiv_org_abs_2409_08845
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle AIPO: Improving Training Objective for Iterative Preference Optimization
Shen, Yaojie
Wang, Xinyao
Niu, Yulei
Zhou, Ying
Tang, Lexin
Zhang, Libo
Chen, Fan
Wen, Longyin
Computation and Language
Preference Optimization (PO), is gaining popularity as an alternative choice of Proximal Policy Optimization (PPO) for aligning Large Language Models (LLMs). Recent research on aligning LLMs iteratively with synthetic or partially synthetic data shows promising results in scaling up PO training for both academic settings and proprietary trained models such as Llama3. Despite its success, our study shows that the length exploitation issue present in PO is even more severe in Iterative Preference Optimization (IPO) due to the iterative nature of the process. In this work, we study iterative preference optimization with synthetic data. We share the findings and analysis along the way of building the iterative preference optimization pipeline. More specifically, we discuss the length exploitation issue during iterative preference optimization and propose our training objective for iterative preference optimization, namely Agreement-aware Iterative Preference Optimization (AIPO). To demonstrate the effectiveness of our method, we conduct comprehensive experiments and achieve state-of-the-art performance on MT-Bench, AlpacaEval 2.0, and Arena-Hard. Our implementation and model checkpoints will be made available at https://github.com/bytedance/AIPO.
title AIPO: Improving Training Objective for Iterative Preference Optimization
topic Computation and Language
url https://arxiv.org/abs/2409.08845