MORepair: Teaching LLMs to Repair Code via Multi-Objective Fine-tuning

Cite Our Work

If you use MORepair in your research, please cite our paper:

@article{yang25morepair,
author = {Yang, Boyang and Tian, Haoye and Ren, Jiadong and Zhang, Hongyu and Klein, Jacques and Bissyande, Tegawende and Le Goues, Claire and Jin, Shunfu},
title = {MORepair: Teaching LLMs to Repair Code via Multi-Objective Fine-Tuning},
year = {2025},
publisher = {Association for Computing Machinery},
issn = {1049-331X},
url = {https://doi.org/10.1145/3735129},
doi = {10.1145/3735129},
journal = {ACM Trans. Softw. Eng. Methodol.},
}

Colab Demo

Explore MORepair with our Colab Notebook: MORepair Demo

📚 Datasets

MORepair is trained on TutorLLMCode, and evaluated on four carefully curated datasets, covering different programming languages and repair scenarios:

🎯 Training Dataset

Dataset	Description	Size	Language	Obtain
TutorLLMCode	High-quality C++ code repair dataset with human and LLM-generated rationales	1.5K	C++	Website

🏆 Evaluation Benchmarks

Dataset	Description	Size	Language	Obtain
EvalRepair-Java	Real-world Java program repair benchmark derived from HumanEval	163	Java	Hugging Face
EvalRepair-C++	Real-world C++ program repair benchmark derived from HumanEval	164	C++	Hugging Face
D4J-Repair	single-function subset of Defects4J	371	Java	Hugging Face
SWE-Repair	single-function subset of SWE-Bench	204	Multi	Hugging Face

💡 Note: All datasets are preprocessed and ready to use. For detailed dataset statistics and usage instructions, please refer to our paper.

MORepair Framework Overview

MORepair is a novel Multi-Objective fine-tuning framework designed specifically for LLM-based program Repair. It steers LLMs toward a precise understanding of the reasoning logic behind the repair process, thereby enabling them to generate high-quality patches.

Key Features

• 🚀 Multi-Objective Fine-Tuning: A novel approach for significantly enhanced code repair capabilities.
• 🧠 Improved Reasoning Logic: Guides LLMs to deeply understand the "why" behind code fixes, not just the "what."
• 🛠️ High-Quality Patch Generation: Empowers LLMs to produce more accurate and reliable code patches.
• 📄 Instruction-Following Enhancement: Particularly effective with instruction-tuned base models.
• 🐳 Dockerized & Reproducible: Easy setup with Docker ensures consistent environments for research and development.
• 🧩 Extensible & Adaptable: Designed to be flexible for various models and custom datasets.

Quick Start & Environment Setup

Get up and running with MORepair using Docker.

1. Prerequisites:

   • docker.io
   • zstd (for decompressing datasets, if you plan to use the provided ones)

2. Build the Docker Image: Clone this repository, then navigate to its root directory and run:

   docker build -t morepair .

3. Run the Docker Container:

   # Mount your local MORepair repository (replace /path/to/your/local/morepair with the actual path)
   docker run -it -v /path/to/your/local/morepair:/opt/morepair morepair
   cd /opt/morepair

   Tip: On Linux/macOS, use $(pwd) for the current path: docker run -it -v $(pwd):/opt/morepair morepair

Using MORepair

Follow these steps to leverage MORepair for your program repair tasks.

1. Prepare Your Custom Dataset

Your fine-tuning dataset should be a JSON file containing a list of dictionaries. Each dictionary must have a single key, text, whose value is a string formed by concatenating: 1. The input (e.g., buggy code with instructions) 2. The output for the first objective (e.g., the rationale or thought process) 3. The output for the second objective (e.g., the corrected code)

These three parts must be separated by an End-Of-Sentence (EOS) token (e.g., </s> for Llama). The entire text value must also end with an EOS token. For instruction-tuned base models, formatting the input as per the model's instruction template is highly recommended for optimal performance.

Refer to data/trainset/llama_llm.json for an example and TutorLLMCode.md for details on the TutorLLMCode dataset structure.

(Optional) Download Preprocessed TutorLLMCode Datasets: Within the Docker container, run:

python3 fetch_data.py

This script downloads llama_human.json (human-generated rationales) and llama_llm.json (GPT-4 generated rationales) for single-file C++ buggy programs.

2. Perform Multi-Objective Fine-tuning

Use the MOTrain.py script for fine-tuning. Key arguments include: * --base_model_name_or_path: Name or path of your base LLM. * --dataset_path: Path to your prepared JSON dataset. * --output_model_dir_name: Subdirectory name under ./models to save the fine-tuned model.

Example (inside Docker):

python3 MOTrain.py \\
    --base_model_name_or_path CodeLlama-7b-Instruct-hf \\
    --dataset_path data/trainset/llama_llm.json \\
    --output_model_dir_name my_custom_codellama7b

3. Model Inference

The fine-tuned model (LoRA adapters and potentially merged model) will be saved in ./models/<your_output_model_dir_name>. For inference, use the model files from the ./codellama_merged (or similarly named) subdirectory.

The inference_cpp.py script provides an example of an inference pipeline. Using 8-bit quantization is recommended to optimize resource usage.

(Optional) Reproducing Paper Results

For those interested in replicating the results from our paper:

1. Full Dataset Download & Setup: Refer to our paper or earlier README.md versions for detailed instructions on acquiring and preparing datasets like the full EvalRepair series, Defects4J, and SWE-bench (often involving *.zst decompression). Example (inside Docker):

   # zstd -d evalrepair-java.zst -o evalrepair-java.tar && tar -xvf evalrepair-java.tar

2. Execution Scripts: Use the provided rqN.sh scripts (e.g., rq1.sh, rq2.sh) within the Docker container. These scripts typically include re-judging options.

3. Specific Model Fine-tuning & Inference: The finetune_and_inference.sh script can be used. Example (inside Docker):

   # python3 fetch_data.py # If TutorLLMCode is not yet downloaded
   # bash finetune_and_inference.sh CodeLlama-13b-Instruct-hf llama_llm codellama13b-stdft 0

   Consult the original README.md's parameter table for model, dataset, and lambda configurations.