Aligning CodeLLMs with Direct Preference Optimization

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Miao, Yibo, Gao, Bofei, Quan, Shanghaoran, Lin, Junyang, Zan, Daoguang, Liu, Jiaheng, Yang, Jian, Liu, Tianyu, Deng, Zhijie
Natura: Preprint
Pubblicazione: 2024
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866909362172198912
author Miao, Yibo
Gao, Bofei
Quan, Shanghaoran
Lin, Junyang
Zan, Daoguang
Liu, Jiaheng
Yang, Jian
Liu, Tianyu
Deng, Zhijie
author_facet Miao, Yibo
Gao, Bofei
Quan, Shanghaoran
Lin, Junyang
Zan, Daoguang
Liu, Jiaheng
Yang, Jian
Liu, Tianyu
Deng, Zhijie
contents The last year has witnessed the rapid progress of large language models (LLMs) across diverse domains. Among them, CodeLLMs have garnered particular attention because they can not only assist in completing various programming tasks but also represent the decision-making and logical reasoning capabilities of LLMs. However, current CodeLLMs mainly focus on pre-training and supervised fine-tuning scenarios, leaving the alignment stage, which is important for post-training LLMs, under-explored. This work first identifies that the commonly used PPO algorithm may be suboptimal for the alignment of CodeLLM because the involved reward rules are routinely coarse-grained and potentially flawed. We then advocate addressing this using the DPO algorithm. Based on only preference data pairs, DPO can render the model rank data automatically, giving rise to a fine-grained rewarding pattern more robust than human intervention. We also contribute a pipeline for collecting preference pairs for DPO on CodeLLMs. Studies show that our method significantly improves the performance of existing CodeLLMs on benchmarks such as MBPP and HumanEval.
format Preprint
id arxiv_https___arxiv_org_abs_2410_18585
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Aligning CodeLLMs with Direct Preference Optimization
Miao, Yibo
Gao, Bofei
Quan, Shanghaoran
Lin, Junyang
Zan, Daoguang
Liu, Jiaheng
Yang, Jian
Liu, Tianyu
Deng, Zhijie
Artificial Intelligence
Machine Learning
The last year has witnessed the rapid progress of large language models (LLMs) across diverse domains. Among them, CodeLLMs have garnered particular attention because they can not only assist in completing various programming tasks but also represent the decision-making and logical reasoning capabilities of LLMs. However, current CodeLLMs mainly focus on pre-training and supervised fine-tuning scenarios, leaving the alignment stage, which is important for post-training LLMs, under-explored. This work first identifies that the commonly used PPO algorithm may be suboptimal for the alignment of CodeLLM because the involved reward rules are routinely coarse-grained and potentially flawed. We then advocate addressing this using the DPO algorithm. Based on only preference data pairs, DPO can render the model rank data automatically, giving rise to a fine-grained rewarding pattern more robust than human intervention. We also contribute a pipeline for collecting preference pairs for DPO on CodeLLMs. Studies show that our method significantly improves the performance of existing CodeLLMs on benchmarks such as MBPP and HumanEval.
title Aligning CodeLLMs with Direct Preference Optimization
topic Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2410.18585