GitHubAD: Bulk Files

GitHubAD pairs 8,401 CVEs with the commits that fixed them, across 2,816 open-source repositories, and freezes the history of each repository's default branch. This Hugging Face dataset holds the bulk files: the commit archives, the precomputed vectors, the keyword scores, and every file that quotes commit text, namely the two files of corrections/ and the full outputs of the language-model runs, 216.3 GB in total. It is the only part of the release that holds commit messages and diffs. The labels, run outputs, shortlist, TREC files, loader, scorer, and documentation are on Zenodo, DOI 10.5281/zenodo.23753278.

Patch tracing is the task of finding, for a disclosed vulnerability, the commit that fixed it. A method receives the CVE description from the National Vulnerability Database and ranks the commits of the affected repository.

Folders

Folder	Files	Size	Contents
commits/	2,816	50.4 GB	one archive per repository, <owner>@@<repo>.tar.zst, with the id, date, message, and diff of every commit reachable from the default branch
corrections/	2	3.4 MB	merge_first_parent_diffs.jsonl.zst, GitHub's first-parent diff of each of the 585 recorded fix commits that are merge commits; malformed_id_map.tsv, the original ids behind the placeholders malformed-1 and malformed-2 in the Zenodo files
runs/	4	11.8 MB	the PatchHolmes agent and IRCoT runs with their model text (reasoning, raw_pred), which quotes commits; Zenodo holds the same runs without it
embeddings/qwen3-embedding-8b/	11,264	126.5 GB	float16 Qwen3-Embedding-8B vectors for 15,394,045 commits and 8,375 CVE queries
embeddings/octen-embedding-8b/	1,362	21.8 GB	float16 Octen-Embedding-8B vectors for the 227 repositories of the 809-CVE working subset
retrieval/bm25_time/	2,816	17.6 GB	keyword scores (BM25 with two time signals), up to 100,000 scored commits per CVE

In 2,815 of the 2,816 archives, the latest commit is dated 10 February 2025 or earlier. The revision tag v2.0.0 marks the files the paper describes; later versions get new tags. SCHEMAS.md on Zenodo gives the record format of every data file here, and metadata.zip on Zenodo lists the SHA-256 checksum of every file in this dataset.

Getting Started

hf download guanqun-yang/GitHubAD --repo-type dataset --revision v2.0.0 --local-dir GitHubAD
# one part only, for example the commit archives:
hf download guanqun-yang/GitHubAD --repo-type dataset --revision v2.0.0 --local-dir GitHubAD --include "commits/*"

Then download the Zenodo files into the same folder and follow the Zenodo README. Its loader tools/load.py reads the labels, commits, vectors, scores, and run outputs, and ranks the commits of a repository on a CPU. Its scorer tools/githubad_score.py scores a run.

Known Limitations

• The archives hold the default branch only, so a fix that exists only on another branch is absent.
• Of the 585 recorded fix commits that are merge commits, 433 have a stored diff that changes a different set of files than GitHub's first-parent diff. Their vectors and keyword scores were computed on the stored diffs. corrections/ holds GitHub's first-parent diffs, and Zenodo holds a subset file of the 8,026 CVEs to report results on next to all CVEs.
• Six repositories have one truncated archive shard each, whose 19,408 readable commits have no vector. No recorded fix commit is among them.
• 107 commits in 37 repositories carry an id that the history extraction mixed with message text. No recorded fix commit is among them.
• 26 CVEs have no released query vector, although Phase 1 ranked them by vector; why their vectors are missing was not determined.
• A keyword-score file keeps at most 100,000 scored commits. In the 12 largest repositories, the files score fewer commits than the history holds, down to 53,232 of 546,234 commits (9.75%) in mjg59/linux. In 11 of 12 randomly drawn repositories, every commit is scored.

Licenses

• Vectors, keyword scores, and documentation: CC BY 4.0.
• Commit messages and diffs: redistributed from public GitHub repositories, each under its repository's own license. metadata/repo_licenses.csv on Zenodo lists the license GitHub reports for every repository, and a subset file lists the 6,222 CVEs of the 2,079 repositories with a recognized license.

Ethics and Removal Requests

All commit data is public on GitHub. A commit record holds the commit id, date, message, and diff, without git's author and committer fields. Names and email addresses can still appear in messages and diffs, for example in Signed-off-by lines. A repository owner or a named person who wants material removed can open a discussion on this page or email the corresponding author. We then remove the messages and diffs of those commits from the next version and from every earlier revision of this dataset, including the tagged v2.0.0. The Zenodo record holds no commit messages or diffs. A removal changes the affected files and their checksums, and the changelog of the next version lists each removal.

Citation

See CITATION.cff on Zenodo. Contact: Guanqun Yang, guanqun.yang@outlook.com.