1. Git clone our repository
2. creating conda environment:

conda create -n LLMenv python=3.10
conda activate LLMenv
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
pip install sympy==1.13.1
pip install --upgrade transformers
conda install psutil
pip install peft
pip install sentencepiece
pip install deepspeed

git clone https://github.com/huggingface/alignment-handbook.git
cd ./alignment-handbook/
python -m pip install .

you can download dataset from https://huggingface.co/datasets/ZhenlongDai/ACPR.

Download it and place it in the directory ./merged_test_cases.

Problem file in "repairDataset/Program_Question_Data/English_Program_Question_StringVersion.json"

file in the directory "repairDataset/RepairData-PythonLevel/CRFLPDataset/"

View "codeTool/use.md"

Please download the pre-trained CodeLlama-7b-Instruct-hf weights from huggingFace and put it in the same directory "./CodeLlama-7b-Instruct-hf".

DATA_FILE="CRFLPDataset"

bash script/pipeline/process_fixbycrflp.sh

The script is executed after specific parameter values are specified based on the result of step 1. DATA_FILE="CRFLPDataset"

bash script/pipeline/process_trace_CRFLP.sh

The script is executed after specific parameter values are specified. DATA_FILE="mixCRFLPDataset"

bash script/pipeline/process_SecondFix.sh

bash script/merge_sft.sh

The script is executed after specific parameter values are specified.

bash script/Prefer/generate_prefer_data.sh

then you should constrcut generation result to training data. you can refer to "./utils/ConstructPerferDataset.py", we screened the data pairs with the highest and lowest consistency greater than 0.1 difference. You also can directly use perference data "dataNew/FixPerferdataset"

PerferDATA_FILE="dataNew/FixPerferdataset"

bash script/pipeline/Prefer/process_Prefer_FixModel.sh

The script is executed after specific parameter values are specified.

bash cript/Execution/Execution_Eval.sh

The script is executed after specific parameter values are specified.

bash script/Postprocessing.sh

you can download dataset from https://huggingface.co/ZhenlongDai/AdaPatcher.

1. base Program Modifier
   the file of weight: "output_dir/loraWeight/fixbycrflp/checkpoint-8000"
   the file of predict result: "predict_dir/loraWeight/fixbycrflp/test-checkpoint-8000.json"
2. bug locator
   the file of weight: "output_dir/loraWeight/trace_CRFLP/checkpoint-14000"
   the file of predict result: "predict_dir/loraWeight/trace_CRFLP/test-checkpoint-14000.json"
3. Program Modifier with Hybrid Training for Selective Reference
   the file of weight: "output_dir/loraWeight/fixbycrflp2/checkpoint-12000"
   the file of predict result: "predict_dir/loraWeight/fixbycrflp2/test-checkpoint-12000.json"
4. merge_sft
   the file of weight: "output_dir/fix_codeLlama"
   
5. Program Modifier with Adaptive Preference Learning
   the file of weight: "output_dir/DpoWeight/DPOP_Fix_ND3V1/checkpoint-1300"
   the file of predict result: "predict_dir/DpoWeight/DPOP_Fix_ND3V1-GEN/test-checkpoint-1300.json"

BibTeX:

@article{dai2025less,
  title={Less is More: Adaptive Program Repair with Bug Localization and Preference Learning},
  author={Dai, Zhenlong and Chen, Bingrui and Zhao, Zhuoluo and Tang, Xiu and Wu, Sai and Yao, Chang and Gao, Zhipeng and Chen, Jingyuan},
  journal={arXiv preprint arXiv:2503.06510},
  year={2025}
}