Saved in:
Bibliographic Details
Main Authors: Yu, Huimu, Wu, Xing, Xu, Haotian, Zhang, Debing, Hu, Songlin
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2410.02229
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916765761536000
author Yu, Huimu
Wu, Xing
Xu, Haotian
Zhang, Debing
Hu, Songlin
author_facet Yu, Huimu
Wu, Xing
Xu, Haotian
Zhang, Debing
Hu, Songlin
contents Large language models (LLMs) have made significant progress in natural language understanding and generation, driven by scalable pretraining and advanced finetuning. However, enhancing reasoning abilities in LLMs, particularly via reinforcement learning from human feedback (RLHF), remains challenging due to the scarcity of high-quality preference data, which is labor-intensive to annotate and crucial for reward model (RM) finetuning. To alleviate this issue, we introduce CodePMP, a scalable preference model pretraining (PMP) pipeline that utilizes a large corpus of synthesized code-preference pairs from publicly available high-quality source code. CodePMP improves RM finetuning efficiency by pretraining preference models on large-scale synthesized code-preference pairs. We evaluate CodePMP on mathematical reasoning tasks (GSM8K, MATH) and logical reasoning tasks (ReClor, LogiQA2.0), consistently showing significant improvements in reasoning performance of LLMs and highlighting the importance of scalable preference model pretraining for efficient reward modeling.
format Preprint
id arxiv_https___arxiv_org_abs_2410_02229
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle CodePMP: Scalable Preference Model Pretraining for Large Language Model Reasoning
Yu, Huimu
Wu, Xing
Xu, Haotian
Zhang, Debing
Hu, Songlin
Artificial Intelligence
Computation and Language
Large language models (LLMs) have made significant progress in natural language understanding and generation, driven by scalable pretraining and advanced finetuning. However, enhancing reasoning abilities in LLMs, particularly via reinforcement learning from human feedback (RLHF), remains challenging due to the scarcity of high-quality preference data, which is labor-intensive to annotate and crucial for reward model (RM) finetuning. To alleviate this issue, we introduce CodePMP, a scalable preference model pretraining (PMP) pipeline that utilizes a large corpus of synthesized code-preference pairs from publicly available high-quality source code. CodePMP improves RM finetuning efficiency by pretraining preference models on large-scale synthesized code-preference pairs. We evaluate CodePMP on mathematical reasoning tasks (GSM8K, MATH) and logical reasoning tasks (ReClor, LogiQA2.0), consistently showing significant improvements in reasoning performance of LLMs and highlighting the importance of scalable preference model pretraining for efficient reward modeling.
title CodePMP: Scalable Preference Model Pretraining for Large Language Model Reasoning
topic Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2410.02229