---
license: cc-by-nc-4.0
language:
- en
task_categories:
- text-generation
tags:
- git
- commit-messages
---

# CommitCraft data

Git diffs paired with Conventional Commit messages, used to fine-tune
`Qwen/Qwen2.5-Coder-1.5B-Instruct` in the CommitCraft project.

## Source and license

Diffs and original messages come from
[CommitBench](https://huggingface.co/datasets/Maxscha/commitbench). This dataset keeps its
**CC BY-NC 4.0** license: non-commercial use only, credit required.

## How it was made

1. CommitBench rows were filtered: diff size, files touched, low-value messages, secrets.
2. `claude-haiku-5-5` wrote a new `message` for each row and gave it a quality score from 1 to 5.
3. Rows were kept only if the message passes the format check and quality is at least 4.
4. Rows were split by project, so no repository appears in two splits.

## Limitations

- `message` was written by a model, so some labels may be wrong.
- CommitBench hides numbers and hashes as `<I>` and `<HASH>` in the diffs.

## Splits

- train: 3,767 rows
- validation: 209 rows
- test: 225 rows