---
license: cc-by-nc-4.0
language:
- en
tags:
- code
size_categories:
- 1M<n<10M
---

# CommitBench: A Benchmark for Commit Message Generation

## EXECUTIVE SUMMARY
We provide CommitBench as an open-source, reproducible and privacy- and license-aware benchmark for commit message generation. The dataset is gathered from GitHub repositories with licenses that permit redistribution. We provide six programming languages, Java, Python, Go, JavaScript, PHP, and Ruby. The commit messages in natural language are restricted to English, as it is the working language in many software development projects. The dataset has 1,664,590 examples that were generated by using extensive quality-focused filtering techniques (e.g., excluding bot commits). Additionally, we provide a version with longer sequences for benchmarking models with more extended sequence input.

## CURATION RATIONALE
We created this dataset due to quality and legal issues with previous commit message generation datasets. Given a git diff displaying code changes between two file versions, the task is to predict the accompanying commit message describing these changes in natural language. We base our GitHub repository selection on that of a previous dataset, CodeSearchNet, but apply a large number of filtering techniques to improve the data quality and eliminate noise. Due to the original repository selection, we are also restricted to the aforementioned programming languages. It was important to us, however, to provide some number of programming languages to accommodate any changes in the task due to the degree of hardware-relatedness of a language. The dataset is provided as a large CSV file containing all samples. We provide the following fields: Diff, Commit Message, Hash, Project, Split.

## DOCUMENTATION FOR SOURCE DATASETS
Repository selection based on CodeSearchNet, which can be found under [https://github.com/github/CodeSearchNet](https://github.com/github/CodeSearchNet).

## LANGUAGE VARIETIES
Since GitHub hosts software projects from all over the world, there is no single uniform variety of English used across all commit messages. This means that phrasing can be regional or subject to influences from the programmer's native language. It also means that different spelling conventions may co-exist and that different terms may be used for the same concept. Any model trained on this data should take these factors into account.

### Overview of split by programming language for CommitBench:
- Java: 153,119
- Ruby: 233,710
- Go: 137,998
- JavaScript: 373,598
- Python: 472,469
- PHP: 294,394

## SPEAKER DEMOGRAPHIC
Due to the extremely diverse (geographically, but also socio-economically) backgrounds of the software development community, there is no single demographic the data comes from. Globally, the average software developer tends to be male and has obtained higher education. Due to the anonymous nature of GitHub profiles, gender distribution information cannot be extracted.

## ANNOTATOR DEMOGRAPHIC
Due to the automated generation of the dataset, no annotators were used.

## SPEECH SITUATION AND CHARACTERISTICS
The public nature and often business-related creation of the data by the original GitHub users fosters a more neutral, information-focused, and formal language. As it is not uncommon for developers to find the writing of commit messages tedious, there can also be commit messages representing the frustration or boredom of the commit author. While our filtering is supposed to catch these types of messages, there can be some instances still in the dataset.

## PREPROCESSING AND DATA FORMATTING
See our paper for all preprocessing steps. We do not provide the un-processed raw data due to privacy concerns, but it can be obtained via CodeSearchNet or requested from the authors.

## CAPTURE QUALITY
While our dataset is completely reproducible at the time of writing, there are external dependencies that could restrict this. If GitHub shuts down and someone with a software project in the dataset deletes their repository, there can be instances that are non-reproducible.

## LIMITATIONS
While our filters are meant to ensure a high quality for each data sample in the dataset, we cannot ensure that only low-quality examples were removed. Similarly, we cannot guarantee that our extensive filtering methods catch all low-quality examples. Some might remain in the dataset. Another limitation of our dataset is the low number of programming languages (there are many more) as well as our focus on English commit messages.

## METADATA
- **License:** Dataset under the CC BY-NC 4.0 license, code under the MIT license

## DISCLOSURES AND ETHICAL REVIEW
While we put substantial effort into removing privacy-sensitive information, our solutions cannot find 100% of such cases. This means that researchers and anyone using the data need to incorporate their own safeguards to effectively reduce the amount of personal information that can be exposed.

## ABOUT THIS DOCUMENT
A data statement is a characterization of a dataset that provides context to allow developers and users to better understand how experimental results might generalize, how software might be appropriately deployed, and what biases might be reflected in systems built on the software.

This data statement was written based on the template for the Data Statements Version 2 schema. The template was prepared by Angelina McMillan-Major, Emily M. Bender, and Batya Friedman and can be found at [https://techpolicylab.uw.edu/data-statements/](https://techpolicylab.uw.edu/data-statements/) and was updated from the community Version 1 Markdown template by Leon Derczynski.