quguanni/kernel-vuln-dataset-full

Linux Kernel Vulnerability-Introducing Commits Dataset

Dataset Description

A labeled dataset of 1,426,202 Linux kernel git commits with full metadata, diffs, and binary labels indicating whether each commit introduced a vulnerability that was later fixed.

Intended use: Training and evaluating models for vulnerability-introducing commit detection — predicting whether a given code change will later require a security or bug fix.

How the Data Was Collected

1. Repository: The full Linux kernel git history was cloned from git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git (all branches and tags, complete history).

2. Metadata extraction: A single git log --all pass extracted structured metadata for every reachable commit (1,426,202 total), including author/committer info, dates, and the full commit message parsed into subject, body, and trailers.

3. Diff extraction: A single git log --all -p --numstat pass extracted the complete unified diff and per-file insertion/deletion counts for every commit.

4. Labeling via Fixes tag mining: The vuln_commits_full.csv dataset maps fixing commits to the commits they fix, identified through the kernel's Fixes: trailer convention. Each commit whose abbreviated hash appears as an introducing_commit in that mapping receives label=1; all others receive label=0.

5. Stratified splitting: The dataset was split 80/10/10 into train/validation/test with stratified sampling (seed=42) to preserve the label distribution across splits.

Column Descriptions

Column	Type	Description
hash	string	Full 40-character commit SHA-1
abbreviated_hash	string	Short commit hash (typically 12 characters)
parent_hashes	string	Space-separated full hashes of parent commits
parent_count	int32	Number of parent commits (0=root, 1=normal, 2+=merge)
author_name	string	Original author's name
author_email	string	Original author's email
author_date	string	Author timestamp in strict ISO 8601 format
committer_name	string	Committer's name (person who applied the patch)
committer_email	string	Committer's email
committer_date	string	Committer timestamp in strict ISO 8601 format
subject	string	First line of the commit message
body	string	Commit message text between subject and trailers
trailers	string	Structured trailer lines (Signed-off-by, Fixes, Reviewed-by, Cc, Link, etc.) joined by newlines
diff_raw	large_string	Complete unified diff output (diff headers, hunks, context lines, etc.)
insertions	int64	Total lines added across all files (from numstat)
deletions	int64	Total lines removed across all files (from numstat)
files_changed	int32	Number of files modified
label	int8	1 = vulnerability-introducing commit, 0 = clean

Split Sizes and Label Distribution

Split	Rows	Label=1	Label=0	Positive %
train	1,140,962	64,002	1,076,960	5.61%
validation	142,620	8,000	134,620	5.61%
test	142,620	8,000	134,620	5.61%
total	1,426,202	80,002	1,346,200	5.61%

Known Limitations

1. Fixes tags capture all bugs, not just security vulnerabilities. The kernel's Fixes: trailer is used for all bug fixes — logic errors, performance regressions, build failures, and similar — not exclusively for security-critical vulnerabilities. The label therefore reflects "introduced a defect that was later fixed" rather than "introduced a security vulnerability" specifically.

2. Label noise from catch-all hashes. Some Fixes: tags point to very early commits (e.g., 1da177e4c3f4 — the initial Linux 2.6.12-rc2 import) when the actual introducing commit is unknown or predates git history. These commits receive label=1 despite not being the true root cause.

3. Right-censoring of recent commits. Commits near the HEAD of the repository have had less time to be identified as bug-introducing. Some recent label=0 commits may in fact contain undiscovered bugs, introducing a systematic false-negative bias toward the end of the timeline.

4. Merge commits have empty diffs. By default, git log -p does not produce diffs for merge commits. Of the 80,002 label=1 commits, 155 (0.19%) are merges with empty diff_raw. These should be excluded or handled specially in modeling.

5. Feature distribution skew. Label=1 commits tend to be larger than label=0 commits (mean diff length 3.8x, mean insertions 8.3x). Models may learn to use commit size as a shortcut rather than understanding code semantics. Consider controlling for commit size in evaluation.

6. Single-project bias. All data comes from one project (the Linux kernel). Models trained on this data may not generalize to other codebases with different coding conventions, review processes, or commit practices.

7. Incomplete Fixes tag coverage. Not all bug-fixing commits in the kernel carry a Fixes: trailer. The labeled set of introducing commits is therefore a lower bound — the true number of bug-introducing commits is likely higher, meaning some label=0 commits are mislabeled.

Citation

If you use this dataset, please cite:

@misc{linux_vuln_commits_2026,
  title={Linux Kernel Vulnerability-Introducing Commits Dataset},
  year={2026},
  note={Derived from the Linux kernel git history and Fixes-tag mining},
  url={https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git}
}

License

The Linux kernel is licensed under GPL-2.0. Commit metadata and diffs are derived from the kernel source repository and are distributed under the same license.