Saved in:
Bibliographic Details
Main Authors: Yang, Haoyan, Le, Khiem, Hua, Ting, Gao, Shangqian, Xu, Binfeng, Tang, Zheng, Xu, Jie, Chawla, Nitesh V., Jin, Hongxia, Srinivasan, Vijay
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2502.05400
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918388393050112
author Yang, Haoyan
Le, Khiem
Hua, Ting
Gao, Shangqian
Xu, Binfeng
Tang, Zheng
Xu, Jie
Chawla, Nitesh V.
Jin, Hongxia
Srinivasan, Vijay
author_facet Yang, Haoyan
Le, Khiem
Hua, Ting
Gao, Shangqian
Xu, Binfeng
Tang, Zheng
Xu, Jie
Chawla, Nitesh V.
Jin, Hongxia
Srinivasan, Vijay
contents Although LLMs have achieved significant success, their reliance on large volumes of human-annotated data has limited their potential for further scaling. In this situation, utilizing self-generated synthetic data has become crucial for fine-tuning LLMs without extensive human annotation. However, current methods often fail to ensure consistent improvements across iterations, with performance stagnating after only minimal updates. To overcome these challenges, we introduce Dynamic Noise Preference Optimization (DNPO), which combines dynamic sample labeling for constructing preference pairs with controlled, trainable noise injection during preference optimization. Our approach effectively prevents stagnation and enables continuous improvement. In experiments with Llama-3.2-3B and Zephyr-7B, DNPO consistently outperforms existing methods across multiple benchmarks. Additionally, with Zephyr-7B, DNPO shows a significant improvement in model-generated data quality, with a 29.4% win-loss rate gap compared to the baseline in GPT-4 evaluations.
format Preprint
id arxiv_https___arxiv_org_abs_2502_05400
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Dynamic Noise Preference Optimization: Self-Improvement of Large Language Models with Self-Synthetic Data
Yang, Haoyan
Le, Khiem
Hua, Ting
Gao, Shangqian
Xu, Binfeng
Tang, Zheng
Xu, Jie
Chawla, Nitesh V.
Jin, Hongxia
Srinivasan, Vijay
Computation and Language
Although LLMs have achieved significant success, their reliance on large volumes of human-annotated data has limited their potential for further scaling. In this situation, utilizing self-generated synthetic data has become crucial for fine-tuning LLMs without extensive human annotation. However, current methods often fail to ensure consistent improvements across iterations, with performance stagnating after only minimal updates. To overcome these challenges, we introduce Dynamic Noise Preference Optimization (DNPO), which combines dynamic sample labeling for constructing preference pairs with controlled, trainable noise injection during preference optimization. Our approach effectively prevents stagnation and enables continuous improvement. In experiments with Llama-3.2-3B and Zephyr-7B, DNPO consistently outperforms existing methods across multiple benchmarks. Additionally, with Zephyr-7B, DNPO shows a significant improvement in model-generated data quality, with a 29.4% win-loss rate gap compared to the baseline in GPT-4 evaluations.
title Dynamic Noise Preference Optimization: Self-Improvement of Large Language Models with Self-Synthetic Data
topic Computation and Language
url https://arxiv.org/abs/2502.05400