GiFT: Gibbs Fine-Tuning for Code Generation

TL;DR

GiFT introduces a novel self-training method for code generation that mitigates bias by sampling from the joint description-code space, improving model performance on challenging benchmarks.

Contribution

The paper proposes Gibbs Fine-Tuning (GiFT), a new self-training approach inspired by Gibbs sampling, to better utilize the joint distribution of descriptions and code in LLM fine-tuning.

Findings

GiFT outperforms baseline models on multiple datasets.

The method effectively addresses long-tail distribution issues.

Empirical results show improved performance on challenging benchmarks.

Abstract

Training Large Language Models (LLMs) with synthetic data is a prevalent practice in code generation. A key approach is self-training, where LLMs are iteratively trained on self-generated correct code snippets. In this case, the self-generated codes are drawn from a conditional distribution, conditioned on a specific seed description. However, the seed description is not the only valid representation that aligns with its intended meaning. With all valid descriptions and codes forming a joint space, codes drawn from the conditional distribution would lead to an underrepresentation of the full description-code space. As such, we propose Gibbs Fine-Tuning (GiFT), a novel self-training method inspired by Gibbs sampling. GiFT allows self-generated data to be drawn from the marginal distribution of the joint space, thereby mitigating the biases inherent in conditional sampling. We provide a theoretical analysis demonstrating the potential benefits of fine-tuning LLMs with code derived from the marginal distribution. Furthermore, we propose a perplexity-based code selection method to mitigate the imbalanced long-tail distribution of the self-generated codes. Empirical evaluation of two LLMs across four datasets demonstrates that GiFT achieves superior performance, particularly on more challenging benchmarks. Source code is available at https://github.com/Alex-HaochenLi/GiFT.