You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
By accessing this dataset you agree to the following terms:
Log in or Sign Up to review the conditions and access this dataset content.
CRSBench Dataset
Benchmark challenge projects for evaluating Cyber Reasoning Systems (CRS).
Description
This dataset contains 134 benchmark challenge projects used by CRSBench to evaluate vulnerability discovery and patch generation capabilities of CRS.
Each challenge project includes:
- Dockerfile and build scripts for reproducible builds
- Ground truth vulnerability metadata (
.aixcc/directory) - Source tarballs (
pkgs/) for the target project
Projects span C/C++ and Java targets including curl, libxml2, SQLite, wireshark, Apache Tika, Jackson Databind, Apache Commons Compress, and more.
Note: The bundled source code has been modified from the original upstream versions to inject test vulnerabilities for CRS evaluation purposes. Original license and copyright files are preserved within each tarball.
Data Use Agreement
By accessing this dataset you agree to:
- No LLM training — Do not use this data to train or fine-tune language models. The dataset contains canary strings for contamination detection.
- Research only — Use this dataset for CRS evaluation, academic research, or benchmarking purposes.
- Attribution — Cite CRSBench and the upstream projects when publishing results. See LICENSE-THIRD-PARTY.md for upstream project licenses.
Canary
This dataset contains canary strings to detect unauthorized use in LLM
training data. Canary identifiers and prefix mappings are tracked in
canary-registry.json in the CRSBench repository.
Usage
Install CRSBench and download:
pip install crsbench
crsbench download --all # all benchmarks with ground truth
crsbench download --all --no-ground-truth # skip .aixcc/ ground truth
crsbench download --benchmark-suite sanity # download a suite
crsbench download --dataset crsbench --benchmarks afc-curl-delta-01 # specific benchmark
Each benchmark is stored as two tarballs:
benchmark.tar.gz— project files, build scripts, and source packagesground-truth.tar.gz—.aixcc/vulnerability metadata and patches
Use --no-ground-truth to download only the project files without
vulnerability answers (useful for blind CRS evaluation).
Dataset Index
This dataset publishes a lightweight index at index/benchmarks.jsonl
(also mirrored as benchmarks-metadata.jsonl at repo root).
The index tracks per-benchmark checksums and metadata so updates are easy to
inspect and diff without extracting tarballs.
Or use huggingface_hub directly:
from huggingface_hub import snapshot_download
# All benchmarks with ground truth
snapshot_download(
repo_id="sslab-gatech/crsbench-dataset",
repo_type="dataset",
local_dir="benchmarks/",
)
# Without ground truth
snapshot_download(
repo_id="sslab-gatech/crsbench-dataset",
repo_type="dataset",
local_dir="benchmarks/",
allow_patterns=["*/benchmark.tar.gz"],
)
Further Reading
- CRSBench repository: https://github.com/sslab-gatech/CRSBench
- Dataset module guide: https://github.com/sslab-gatech/CRSBench/blob/main/docs/modules/dataset.md
- Benchmark developer guide: https://github.com/sslab-gatech/CRSBench/blob/main/docs/benchmark-developer-guide.md
- Data protection and contamination design: https://github.com/sslab-gatech/CRSBench/blob/main/docs/design/benchmark-protection-and-contamination.md
License
CRSBench is licensed under MIT. Bundled upstream source code retains its original license — see LICENSE-THIRD-PARTY.md in this repository or the source version.
- Downloads last month
- 321