NprRF

We provide the necessary scripts and data for training.

|--Generator: patch generator training and evaluation code
​	|--CodeBert
​	|--CodeT5
​	|--UniXcoder
|--Reward: patch assessment model training and evaluation code
​	|--APPT
​	|--Metric (CodeBLEU+BERTScore)
|--RL: implementation of reinforcement learning algorithm
​	|--PPO
​	|--REINFORCE
|--Defects4J_method_singlehunk.csv: list of method level buggy projects in Defects4J
|--Defects4J_method_singlehunk.json: list of method level buggy projects in Defects4J
|--CodeQL.sh: vulnerability detection script
|--RL_run_metric_REINFORCE.sh
|--RL_run_REINFORCE.sh
|--RL_run_metric.sh
|--RL_run.sh
|--Defects4J_30_测试结果.xlsx: Defects4J test results

Dataset

The raw dataset can be downloaded from this link.

Environment

• python 3.8
• numpy==1.24.3
• pandas==2.0.3
• scikit_learn==1.3.0
• torch==2.4.1+cu121
• transformers==4.42.3

Defects4J: please configure the Defects4J environment yourself.

Train & Evaluate

Model_type: CodeBert / CodeT5 / UniXcoder

Reward_type: APPT / Metric

Train

1. Repair Training

• Please modify the Generator/Model_type/train.sh, this file has some parameters needed to train our model.

  bash Generator/Model_type/train.sh

2. Reinforcement Learning

• Please modify the RL_run.sh (for APPT) or RL_run_metric.sh (for Metric), this file has some parameters needed to train our model.

  bash RL_run.sh
  bash RL_run_metric.sh

Evaluate (for Defects4J)

• Please modify the Generator/Model_type/Defects4J_experiment.sh, reset the model loading path and some parameters required for the model.

bash Generator/Model_type/Defects4J_experiment.sh