RepairLLaMA: Efficient Representations and Fine-Tuned Adapters for Program Repair

If you use RepairLLaMA in academic research, please cite "RepairLLaMA: Efficient Representations and Fine-Tuned Adapters for Program Repair", IEEE Transactions on Software Engineering, 2025.

@article{repairllama2023,
  title={RepairLLaMA: Efficient Representations and Fine-Tuned Adapters for Program Repair},
  author={Silva, Andr{\'e} and Fang, Sen and Monperrus, Martin},
  journal = {IEEE Transactions on Software Engineering},
  doi = {10.1109/TSE.2025.3581062},
  url = {http://arxiv.org/abs/2312.15698}
}

This repository contains the code, model, and results to replicate the paper "RepairLLaMA: Efficient Representations and Fine-Tuned Adapters for Program Repair"

It is structured as follows:

• repairllama-lora contains the RepairLLaMA low-rank adaptation of CodeLLaMA-7B, called "repair adapter"
• results contains all generated patches for Defects4J and HumanEval-Java by all models (incl. full fine-tuning, lora, and code representations)
• src contains the training and inference scripts, and scripts to generate datasets for different input-output representations (IRxOR)
• example contains an example notebook explaining how to load and prompt the RepairLLaMA model
• benchmarks contains the datasets for different input-output representations (IRxOR)

Models

All fine-tuned models are available on HuggingFace, here are specific links:

• IR1xOR1: https://huggingface.co/ASSERT-KTH/RepairLLaMA-IR1-OR1
• IR1xOR3: https://huggingface.co/ASSERT-KTH/RepairLLaMA-IR1-OR3
• IR1xOR3: https://huggingface.co/ASSERT-KTH/RepairLLaMA-IR1-OR4
• IR2xOR2: https://huggingface.co/ASSERT-KTH/RepairLLaMA-IR2-OR2
• IR3xOR2: https://huggingface.co/ASSERT-KTH/RepairLLaMA-IR3-OR2

Datasets

The processed fine-tuning datasets are made available on HuggingFace at https://huggingface.co/datasets/ASSERT-KTH/repairllama-datasets. It contains the datasets used for training the RepairLLaMA models, one subset per input/output representation pair. To get the 30k..50k datasets we did further filtering based on the token length of input + output pairs being less than 1024 tokens.

If it interests you, you can also find these on our HuggingFace org:

• Megadiff (original dataset, in HF format): https://huggingface.co/datasets/ASSERT-KTH/megadiff
• Megadiff Single-Function (single-function diffs only, with buggy and fixed functions extracted from it): https://huggingface.co/datasets/ASSERT-KTH/megadiff-single-function

Benchmarks

The evaluation benchmarks are Defects4J v2, HumanEval-Java, and GitBug-Java.

We focus on single-function bugs (i.e. bugs whose developer patch exclusively changes one function):

• Defects4J contains 488 single-function bugs: defects4j_sf.txt
• HumanEval-Java contains 162 single-function bugs: humanevaljava_sf.txt
• GitBug-Java contains 90 single-functions bugs: gitbugjava_sf.txt

Note that the original HumanEval-Java contains a duplicate bug.