ajsbsd/openbsd-commits-alpaca

OpenBSD Commit History — Alpaca Format (v1)

Fine-tuning dataset derived from the full commit history of the OpenBSD src repository, structured for instruction fine-tuning in Alpaca format.

Task: given a unified diff, generate the commit message.

---

Dataset Summary

Field	Value
Examples	103,383
Size	~202 MB
Format	Alpaca JSONL
Date range	2000-01-01 → present
Source	openbsd/src (GitHub mirror)
License	ISC

---

Format

Each line is a JSON object:

{
  "system": "You are an expert OpenBSD systems programmer. Given a git diff, write a concise, accurate commit message in the style of the OpenBSD project: terse, imperative, no fluff.",
  "instruction": "Write a commit message for the following diff:\n\n```diff\n--- a/sys/kern/uipc_socket.c\n+++ b/sys/kern/uipc_socket.c\n@@...\n```",
  "input": "",
  "output": "fix use-after-free in socket teardown path"
}

Compatible with Axolotl, Unsloth, and any trainer that accepts Alpaca-format JSONL.

---

Source & Filtering

Extracted from the full git history of openbsd/src using a single streaming git log -p pass. Commits were kept if they met all of:

• Commit message between 10 and 4,000 characters
• Diff between 2 lines and 8,192 characters (2k tokens)
• Not a merge commit
• Not a mass-import or CVS sync commit
• Not touching only generated/lock files

Skip breakdown (v1)

Reason	Count
diff_too_small	68,431
msg_too_long	46,254
diff_too_large	1,441
msg_too_short	491
skip_prefix	154

The 46k msg_too_long commits are a known gap — many contain extended rationale, reviewer sign-offs, and CVS metadata that push past the 4,000 character ceiling. A v2 will handle these with truncation strategies.

---

Skip Breakdown (v2)

Reason	Count
diff_too_large	22,866
msg_too_short	11,656
skip_prefix	1,829
diff_too_small	1,329
msg_too_long	9

💡 Parser Fix (v2 Update): The massive msg_too_long count seen in v1 (46,254 skips) was identified as a streaming parser logic bug that accidentally swapped the commit message and the diff payloads. In v2, the streaming logic has been hardened to split cleanly on native \ndiff --git boundaries. This successfully recovered 183,159 high-quality training examples, dropping msg_too_long anomalies down to just 9 total instances. Remaining skips are now purely reflective of true context limits (diff_too_large) and trivial cleanups (msg_too_short).

---

Reproducing

git clone https://github.com/openbsd/src.git openbsd-src
python3 extract_openbsd.py --repo ./openbsd-src --output openbsd_commits.jsonl

The extraction script is included in this repository as extract_openbsd.py.

---

Intended Use

• Fine-tuning LLMs to write terse, high-quality commit messages
• Teaching models systems programming idioms (C, kernel, crypto, networking)
• Studying the OpenBSD project's engineering culture and code review style

OpenBSD is known for rigorous security auditing, minimal dependencies, and extremely disciplined commit hygiene — making this a high signal-to-noise corpus compared to typical open source repositories.

---

Not Intended For

• Training models to produce insecure or low-quality C code
• Any use that conflicts with the ISC license terms of the source material

---

Citation

If you use this dataset, please credit the OpenBSD project and link to this dataset page.

@misc{openbsd-commits-alpaca-v1,
  title  = {OpenBSD Commit History, Alpaca Format (v1)},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ajsbsd/openbsd-commits-alpaca},
  note   = {Derived from the OpenBSD src repository under ISC license}
}