Artifacts of the paper under review by ESEC/FSE

This repository has been deprecated. Please refer to this link for the latest version.

--------------------------------------------------------------------

This is the online repository of CCT5: A Code-Change-Oriented Pre-Trained Model, a research paper under review by ESEC/FSE. We release the source code and relevant data of CCT5, the data used in our evaluation, as well as the experiment results.

• Getting Started

pytorch==1.8.0
cudatoolkit=11.1
datasets==1.18.3
transformers==4.16.2
tensorboard==2.8.0
tree-sitter==0.19.1

• Dataset

We provide the datasets of pretraining and three downstream tasks. The datasets should be downloaded and uncompressed in the data directory.

pretraining/CodeChangeNet.tar.lrz contains the dataset used in pretraining, i.e., CodeChangeNet. CodeChangeNet is a collection of over 1000 star projects written in six popular programming languages: Go, Java, JavaScript, PHP, Python, and Ruby.

finetune/MessageGeneration contains the download and process script of task1 - Commit Message Generation;

finetune/CommentUpdate contains the dataset of downstream task2 - Just-in-Time Comment Update;

finetune/JITDefectPrediction contains the dataset of downstream task3 - Just-in-Time Defect Prediction;

• Pretrain the model

cd sh
bash pretrain.sh

• Finetune and evaluate the downstream task

Commit Message Generation

cd sh
bash finetune_msggen.sh

Just-in-Time Comment Update

cd sh
bash finetune_cup.sh

Just-in-Time Defect Prediction

cd sh  
bash finetune_jit.sh

• Results

The experiment results of two generation tasks and the ablation study are stored in results with the directory MessageGeneration, CommentUpdate, and Ablation, respectively.