README.md

---
dataset_info:
  features:
  - name: diff
    dtype: string
  - name: message
    dtype: string
  - name: reasoning_trace
    dtype: 'null'
  - name: repo
    dtype: string
  - name: license
    dtype: string
  - name: language
    dtype: string
  splits:
  - name: train
    num_bytes: 57994147
    num_examples: 52173
  - name: validation
    num_bytes: 3248866
    num_examples: 2898
  - name: test
    num_bytes: 3352948
    num_examples: 2898
  download_size: 29457578
  dataset_size: 64595961
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
tags:
- code
- commit-messages
- conventional-commits
- git
- diffs
- source-code
- fine-tuning
size_categories:
- 10K<n<100K
task_categories:
- text-generation
license: other
language:
- en
- code
---

# Committed — Conventional Commits dataset

Filtered (diff -> Conventional Commits message) pairs derived from
[CommitChronicle](https://huggingface.co/datasets/JetBrains-Research/commit-chronicle),
for fine-tuning small models to write commit messages from a diff. Built by the
[Committed](https://github.com/marzoukbaig14/Committed) project.

## Schema

| Field | Type | Notes |
|---|---|---|
| `diff` | string | The code diff for a single-file change. |
| `message` | string | Normalized Conventional Commits subject line (the training target). |
| `reasoning_trace` | string \| null | Reserved for v2 (reasoning distillation); always `null` here. |
| `repo` | string | Source repository (provenance). |
| `license` | string | Source repository's license (provenance). |
| `language` | string | Programming language, identified by file extension. |

## Composition

| Split | Rows |
|---|---|
| train | 52,173 |
| validation | 2,898 |
| test | 2,898 |

**Languages** (identified by file extension):

| Language | Rows | % |
|---|---|---|
| TypeScript | 6,000 | 10.4% |
| JavaScript | 6,000 | 10.4% |
| Python | 6,000 | 10.4% |
| Go | 6,000 | 10.4% |
| Java | 6,000 | 10.4% |
| Rust | 6,000 | 10.4% |
| Shell | 4,215 | 7.3% |
| C++ | 3,753 | 6.5% |
| PHP | 2,708 | 4.7% |
| C | 2,407 | 4.2% |
| C# | 2,146 | 3.7% |
| Swift | 2,129 | 3.7% |
| Kotlin | 1,812 | 3.1% |
| Dart | 1,370 | 2.4% |
| Ruby | 750 | 1.3% |
| Elixir | 679 | 1.2% |

**Commit types:**

| Type | Rows | % |
|---|---|---|
| fix | 28,366 | 48.9% |
| feat | 7,706 | 13.3% |
| chore | 5,959 | 10.3% |
| test | 5,214 | 9.0% |
| refactor | 5,055 | 8.7% |
| docs | 2,482 | 4.3% |
| ci | 1,336 | 2.3% |
| style | 868 | 1.5% |
| build | 562 | 1.0% |
| perf | 421 | 0.7% |

## How it was built

Starting from CommitChronicle, a commit is kept only if:

- the subject line matches a relaxed Conventional Commits pattern
  (`feat|fix|refactor|docs|test|chore|perf|style|build|ci`, optional scope,
  optional breaking `!`), then normalized (lowercase type, `doc` -> `docs`, strip
  `!`, subject line only, trim, strip one trailing period);
- the subject is 5-200 characters;
- it touches exactly one file, and that file is a recognized **code** file by
  extension (the per-repo language attribute is ignored because it mislabels
  polyglot repos);
- it is not a merge, revert, or bot commit (e.g. Dependabot, detected by message
  pattern);
- the diff is at most 2048 tokens (Qwen3-1.7B tokenizer); over-cap diffs are
  dropped, not truncated.

The pool is then balanced (each language capped to 6,000 rows, languages with
fewer than 500 rows dropped) and split 90/5/5 train/validation/test, **stratified
by commit type** so each split preserves the type distribution.

## Provenance & license

Each row keeps its source `repo` and `license`. CommitChronicle aggregates
permissively-licensed repositories (MIT, Apache-2.0, BSD-3-Clause); this
derivative is redistributed under those source terms. Please cite CommitChronicle
and its paper:

> Eliseeva et al., *From Commit Message Generation to History-Aware Commit Message Generation*, arXiv:2308.07655.

## Known limitations

- The source scan covered ~85-90% of CommitChronicle's train split, not a full
  pass, so the language mix is near-complete rather than exhaustive.
- Commit types are imbalanced (`fix` is the plurality); a trivial
  always-predict-`fix` baseline scores around its share, so read prefix-accuracy
  against that floor.
- Description casing is not normalized (acronyms are preserved) — an accepted v1
  limitation.
- No automated scrubbing of secrets/PII; the sensitive-data caveat from
  CommitChronicle is carried forward.