SyzFix collects fixed Linux kernel bugs from syzbot and assembles a structured dataset capturing the full bug-fix lifecycle — from the initial crash report through patch iterations and reviewer discussions to the final merged commit.
Intended for fine-tuning language models to generate and review kernel patches, for researching patch evolution patterns in the Linux development process, and for studying cross-layer kernel bugs — bugs where the crash occurs in one architectural layer but the fix belongs in another (e.g., a specific filesystem crash fixed in the VFS layer).
Website & Documentation: https://sysec-uic.github.io/SyzFix/
Code: https://github.com/sysec-uic/SyzFix
Dataset: https://huggingface.co/datasets/xiaoguangwang/syzfix-dataset
git clone https://github.com/sysec-uic/syzfix.git
cd syzfix
python3 -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -e .The editable install exposes two packages, dataset (collection pipeline,
viewer, HuggingFace sync) and analysis (heuristic analyzers), so downstream
projects can pip install -e path/to/syzfix and import them directly.
# Download and unpack the full per-bug JSON files (~2GB download, ~11GB unpacked)
python -m dataset.restore_processed --repo xiaoguangwang/syzfix-dataset
# Build a lightweight index (~4 MB) instead of loading all 11 GB of processed data on every data viewer call
python -m dataset.view build-index
# List all bugs — V=patch versions, P=has patch, R=has C reproducer, D=has discussion
python -m dataset.view list
# Run all analyzers and show the results
python -m analysis.run_all
python -m analysis.run_all --showData lives in dataset/data/ by default; set SYZFIX_DATA_DIR to point at a
data directory elsewhere (and SYZFIX_RESULTS_DIR for saved analyzer results).
For full details on the dataset, command references, data sources, project structure, and in-depth cross-layer analysis, please visit our official website:
👉 https://sysec-uic.github.io/SyzFix/
Code is released under the MIT License. The dataset aggregates publicly available content from syzbot, lore.kernel.org, and git.kernel.org; see the HuggingFace dataset card for details.