Aligning CodeLLMs with Direct Preference Optimization
This work addresses alignment challenges for CodeLLMs, which is crucial for enhancing their decision-making and reasoning capabilities in programming tasks, representing an incremental improvement over existing methods.
The authors tackled the under-explored alignment stage for CodeLLMs by identifying issues with the PPO algorithm and advocating for the DPO algorithm, which significantly improved performance on benchmarks like MBPP and HumanEval.
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.