Automated Program Repair using Large Language Models (LLMs)

Abstract

This work proposes an approach to APR by integrating a fine-tuned CodeLLaMA-7b model with a GraphRAG-based retrieval framework. We first applied LoRA to fine-tune the CodeLLaMA model using the domain-specific RepairLLaMA dataset. To enhance contextual awareness, we built a graph-based retriever that combines CodeBERT-generated embeddings with structural relationships between buggy and fixed code pairs from the Defects4J dataset. Our system prioritizes relevant repair examples and enables efficient retrieval of code context. Additionally, we developed a simple web interface to provide real-time bug fix suggestions, demonstrating the practical applicability of our pipeline.