README.md

---
license: other
license_name: various-permissive-open-source-licenses
task_categories:
- text-generation
language:
- en
tags:
- commit-message
- conventional-commits
- git
- software-engineering
- code
size_category:
- 100K<n<1M
pretty_name: Commit Messages from High-Quality Repositories
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.jsonl
  - split: validation
    path: data/val.jsonl
  - split: test
    path: data/test.jsonl
- config_name: training_view
  data_files:
  - split: train
    path: training_view/train.jsonl
  - split: validation
    path: training_view/val.jsonl
  - split: test
    path: training_view/test.jsonl
---

# Commit Messages from High-Quality Repositories

292,269 cleaned git commit messages scraped from the full histories of 15 well-regarded open-source
projects, balanced across two styles: **normal** (196,372) and
**conventional commits** (95,897).

## Dataset Summary

- Each record contains the commit `subject`, `body`, plus metadata: `repo`, `sha`, `date`,
  `author`, and labels: `style` (normal/conventional), `type` (fix, feat, docs, ...),
  `scope`, `breaking`.
- Heavy cleaning: GitHub squash suffixes `(#123)` stripped, bot/automated commits removed,
  canned boilerplate removed, exact duplicates removed, git trailers (Signed-off-by etc.)
  stripped from bodies, junk subjects ("wip", "update", ...) removed.
- **Temporal split**: train is the oldest ~98% of commits, validation the next ~1%, test the
  newest ~1%. No train/eval temporal leakage: every eval commit is newer than every train commit.

## Supported Tasks

- `text-generation`: commit message generation, style modeling, conventional-commit
  classification, software-engineering NLP.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("USERNAME/REPO", "default")          # full metadata records
view = load_dataset("USERNAME/REPO", "training_view")  # prompt/completion pairs (body -> subject)
```

`training_view` maps a change description (the commit body) to the commit subject line and is
formatted for fine-tuning tools such as Unsloth, with explicit `prompt` and `completion` columns.

## Source Repositories

| Repository | URL | Commits in dataset |
|---|---|---|
| postgres | https://github.com/postgres/postgres | 57,971 |
| openssl | https://github.com/openssl/openssl | 53,169 |
| angular | https://github.com/angular/angular | 39,465 |
| curl | https://github.com/curl/curl | 37,130 |
| ant-design | https://github.com/ant-design/ant-design | 22,853 |
| neovim | https://github.com/neovim/neovim | 22,096 |
| babel | https://github.com/babel/babel | 13,719 |
| vite | https://github.com/vitejs/vite | 8,209 |
| eslint | https://github.com/eslint/eslint | 7,837 |
| tmux | https://github.com/tmux/tmux | 7,618 |
| vue-core | https://github.com/vuejs/core | 7,540 |
| vitest | https://github.com/vitest-dev/vitest | 5,120 |
| nestjs | https://github.com/nestjs/nest | 4,147 |
| chart-js | https://github.com/chartjs/Chart.js | 3,718 |
| axios | https://github.com/axios/axios | 1,677 |

## Upstream Licenses

Commit messages are inherited from their source projects. Repositories under copyleft or
source-available licenses (git: GPL-2.0; redis: BSD-3 / RSALv2 / SSPLv1) were **excluded**
as a precaution: software licenses technically cover the licensed program rather than VCS
metadata, but commit messages are therefore not openly licensed either, and longer commit
bodies can constitute original expression of their authors. Only permissively-licensed
projects are included:

| Repository | License |
|---|---|
| angular/angular | MIT |
| vuejs/core | MIT |
| vitejs/vite | MIT |
| babel/babel | MIT |
| eslint/eslint | MIT |
| ant-design/ant-design | MIT |
| axios/axios | MIT |
| nestjs/nest | MIT |
| chartjs/Chart.js | MIT |
| vitest-dev/vitest | MIT |
| curl/curl | curl (MIT-style) |
| neovim/neovim | Apache-2.0 / VIM |
| tmux/tmux | ISC |
| openssl/openssl | Apache-2.0 |
| postgres/postgres | PostgreSQL License |

Users should still respect upstream license terms when redistributing or using this dataset commercially.

## Cleaning Statistics

| Filter | Records dropped |
|---|---|
| duplicate | 77,038 |
| bad_subject | 16,190 |
| bot_or_automated | 26,432 |
| license_excluded | 83,369 |
| canned_message | 9,943 |

## Splits

| Split | Records | Date range |
|---|---|---|
| train | 286,425 | 1996-07-09 → 2026-07-14 |
| validation | 2,922 | 2026-07-14 → 2026-08-20 |
| test | 2,922 | 2026-08-20 → 2026-09-23 |

## Privacy

- No author names or email addresses are included. Records are identified only by content
  `sha` + repository, which preserves verifiable provenance without identifying individuals.
- Note: commit bodies may reference contributor names in prose (e.g. credit lines), as these
  are part of the historical commit text itself.

## Limitations

- English-dominant; commit messages from a single ecosystem of C/C++ and JavaScript/TypeScript projects.
- No diffs attached: the dataset describes *what was written*, not the code change itself.
- Bodies are truncated at 2,000 characters.
- Human authors made the source commits; quality varies by repository and era.

## Reproduction

Built with a two-stage pipeline (blobless `git clone --filter=blob:none`, then
`git log` export and rule-based cleaning). See `stats.json` for exact counts.