Aligning CodeLLMs with Direct Preference Optimization

TL;DR

This paper introduces a novel alignment method for CodeLLMs using the DPO algorithm, which leverages preference data pairs to improve model performance on programming benchmarks, surpassing traditional PPO approaches.

Contribution

The work proposes using DPO for CodeLLM alignment, providing a new pipeline for collecting preference data and demonstrating significant performance improvements.

Findings

DPO outperforms PPO in CodeLLM alignment tasks.

The proposed method improves performance on MBPP and HumanEval benchmarks.

A new pipeline for collecting preference pairs enhances alignment quality.

Abstract

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.