Elib27/conventional-commits

Git Diff → Conventional Commit Messages

A dataset of 12,433 (git diff, commit message) pairs scraped from real open-source TypeScript repositories, filtered for quality and formatted for fine-tuning small language models to generate Conventional Commits.

Built as part of a project to fine-tune a local LLM to write commit messages as a prepare-commit-msg git hook. Full write-up: eliotbas.com/projects/commits-fine-tuning

Dataset details

Source repositories

13 open-source TypeScript repositories selected for commit discipline and diversity:

• Libraries: Vite, Vitest, Vue.js core, Angular, tRPC, TanStack Query
• Apps/platforms: Supabase, Nuxt, Twenty, Prisma, TypeORM
• Tooling: typescript-eslint, commitlint
• Workflow-specific: Botpress OSS (to match the author's internship codebase)

Format

JSONL, chat format. Each example has a messages field with two turns:

{
  "messages": [
    {
      "role": "user",
      "content": "diff --git a/src/index.ts b/src/index.ts\n..."
    },
    {
      "role": "assistant",
      "content": "feat(core): add retry logic for failed requests"
    }
  ],
  "meta": {
    "repo": "vitejs/vite",
    "hash": "496a63ad5af3b9e5a7bba5477057f1e61a267a70"
  }
}

Filtering and quality criteria

Not every commit was kept. The main filters applied:

• Conventional Commits enforcement — message must match type(scope)?: description format
• Trailing ref stripping — (#1234) references removed; not inferable from the diff and would be learned as hallucination fodder
• Generated-file exclusion — each commit must touch at least one non-generated file (lockfiles, minified JS, dist/, snapshots excluded)
• Per-repo cap — prevents large repos (e.g. Angular) from dominating the distribution
• Length filter — too small and too big commits are filtered

Commit type distribution

Type	Share
fix	~30%
chore	~18%
docs	~14%
feat	~13%
test	~8%
refactor	~7%
ci / build / perf / style	~10% combined

Key statistics

Metric	Value
Commit message length (p50)	54 chars
Commit message length (p95)	91 chars
Diff length (p50)	2,024 chars
Diff length (p95)	7,827 chars

Intended use

Designed for LoRA fine-tuning of small code language models (0.5B–3B parameters) on the task of generating a single Conventional Commit message from a git diff --staged output.

Models fine-tuned on this dataset were evaluated in the accompanying article. Best results with:

• Qwen2.5-Coder-0.5B-Instruct
• Qwen2.5-Coder-1.5B-Instruct
• Qwen2.5-Coder-3B-Instruct

Limitations

• TypeScript-specific: sourced entirely from TypeScript repos; may not generalize to other languages or mixed codebases
• High-discipline repos only: all source repos follow Conventional Commits strictly; behavior on messier codebases is untested
• Diff-only context: the model sees only the diff, not the ticket, PR description, or broader codebase context — the human author always had more information

Related resources

• Article: eliotbas.com/projects/commits-fine-tuning
• GitHub (scraping scripts, training and evaluation code, git hook): github.com/Elib27/commits-fine-tune
• Fine-tuned models: (HuggingFace model links)