Coeditor is a transformer model that auto-edits your code based on your recent code changes. This repo includes the server code for the Coeditor VSCode extension and the scripts for data processing, model training, and evaluation. The ideas behind Coeditor are presented in the ICLR Spotlight paper, Coeditor: Leveraging Repo-level Diffs for Code Auto-editing, by Jiayi Wei, Greg Durrett, and Isil Dillig.

Watch the Coeditor demo on Youtube.

This project uses poetry to manage the package dependencies. Poetry records all dependencies in the pyproject.toml file and manages the (project-specific) virtual environment for you.

You can install poetry via the following command:

curl -sSL https://install.python-poetry.org | python3 -
poetry completions bash >> ~/.bash_completion

To install all dependencies required by Coeditor, make sure you have python 3.11 installed, then, run the following at the project root:

poetry install

You can then spawn a shell within the project's virtual environment via poetry shell.

Alternatively, you can also install all dependencies using the exported requirements.txt file.

pip3 install -r requirements.txt

Note: All scripts below should be run within the poetry shell (or the virtual environment in which you installed all the dependencies).

Run python scripts/start_server.py to start the Coeditor VSCode extension server. This will download the pre-trained Coeditor model from Huggingface (if not already) and start the extension service at port 5042.

• As an alternative to using the VSCode extension, you can directly run Coeditor inside this notebook by specifying a target file and line nubmer.

You can run all unit tests via poetry run pytest (or just pytest if you run inside the poetry shell).

1. (Optional) Configure the directories. Create the file config/coeditor.json and use the following template to specify where you want to store the dataset and the trained models:

{
    "datasets_root": "/path/to/datasets/directory",
    "models_root": "/path/to/models/direcotry"
}

2. Run the cells in notebooks/download_data.ipynb to clone the repos from GitHub. Note that we use the GitHub search API to search for repos with permissive licenses, so the results may change over time even though the query remains the same.

3. (Optional) Run scripts/prepare_data.py to preprocess the repos into the PyCommits format introduced in the paper. You can safely skip this step since it will automatically be run when you train a new model (and with the corresponding encoder parameters).

Use the scripts/train_model.py script to train a new model from scratch. By default, this script trains a model under our default settings, but you can uncomment the corresponding function calls at the bottom of the script to train a model following one of the ablation settings in the paper.

Note: Only training with a single GPU is tested. You can set the GPU to use via the CUDA_VISIBLE_DEVICES environment variable.

• Comparison with Code Completion Approaches: Run scripts/code_completion_eval.py to obtain the results reported in section 4.1 of the paper.
• Multi-round editing: Run scripts/multi_round_eval.py to obtain the results reported in section 4.2 of the paper.
• Ablation Studies: Run scripts/single_round_eval.py to obtain the results reported in section 4.3 of the paper.

Please cite our paper as:

@inproceedings{
wei2024coeditor,
title={Coeditor: Leveraging Repo-level Diffs for Code Auto-editing},
author={Jiayi Wei and Greg Durrett and Isil Dillig},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=ALVwQjZRS8}
}