README.md

metadata

pretty_name: CheckerBench
language:
  - en
size_categories:
  - n<1K
tags:
  - code
  - static-analysis
  - vulnerability-detection
  - codeql
  - clang-static-analyzer
  - benchmark
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test.parquet

CheckerBench Dataset

300 static-analysis tasks with complete runtime assets: 159 CSA tasks and 141 CodeQL tasks.

Backend / language	Tasks
CSA (C/C++)	159
CodeQL Go	30
CodeQL Java	30
CodeQL JavaScript	30
CodeQL Python	51
Total	300

dataset/tasks.jsonl is the authoritative portable roster. tasks/ contains browsable task inputs. The assets/ directory provides complete task bundles, reference implementations, source snapshots, CodeQL databases, CSA build assets, refinement environments, and a frozen Docker image.

Load the task table

from datasets import load_dataset

tasks = load_dataset("Benchanything/checkerbench", split="test")
assert len(tasks) == 300

The single test split contains all 300 benchmark tasks; no training or validation split is provided. Each row includes the task ID, static analyzer (static_analyzer: csa or codeql), language, CVE and source-project identifiers, task prompt, patch, before/after context, and task configuration as JSON text. checker_template contains the CSA starter implementation and is null for CodeQL tasks. The original browsable inputs are also available under tasks/. task_path identifies the task inside its restored bundle, while environment_id identifies its runtime environment.

Download and restore

Requirements: Linux x86-64, Docker, Python 3.10+, and at least 100 GB of available disk space. Additional task workspaces require additional storage.

Download the complete distribution, including the datasets and frozen runtime image (approximately 13.3 GB of compressed assets):

pip install -U huggingface_hub
hf download Benchanything/checkerbench --repo-type dataset --local-dir checkerbench
cd checkerbench
python3 scripts/restore.py --assets assets --data runtime_data --load-image

For a private repository, authenticate with hf auth login using an account with access before downloading. Public downloads do not require authentication.

To download only dataset archives using a local copy of this repository's scripts:

python3 scripts/download.py --kind dataset --output assets

The downloader verifies every part against its SHA-256 checksum, and the restore script also verifies the complete compressed stream. Split archives are extracted or loaded directly without writing an extra combined copy. Verified downloads and completed restores can be reused.

Create a task workspace

The checkerbench-runtime:review-v1 image targets linux/amd64 and includes LLVM/Clang 18.1.8, CodeQL 2.26.4, language toolchains, and Python dependencies. Model-service credentials are supplied separately by the operator.

Start a container with stable paths:

mkdir -p work
docker run --rm -it --network none --entrypoint /bin/bash \
  -v "$PWD:/release:ro" -v "$PWD/runtime_data:/data:ro" \
  -v "$PWD/work:/work" checkerbench-runtime:review-v1

Prepare a CSA task:

python /release/scripts/prepare.py csa pair_134943 \
  --data-root /data --output /work/pair_134943
cd /work/pair_134943
# Write checker.cpp, then compile and scan.
bash compile.sh
CHECKER_NAME=YOUR_CHECKER_NAME bash scan/scan_before.sh
CHECKER_NAME=YOUR_CHECKER_NAME bash scan/scan_after.sh

Prepare a CodeQL task:

python /release/scripts/prepare.py codeql morefixes-32ed35a8bd8aba7b47ad \
  --language go --data-root /data --output /work/go-example
cd /work/go-example/tasks/morefixes-32ed35a8bd8aba7b47ad
# Edit query/vulnerability.ql, then compile and scan.
bash scripts/compile_query.sh
bash scripts/run_before.sh
bash scripts/run_after.sh

Only tasks in the fixed 300-task roster are accepted. Maintainers can add --reference to reproduce a bundled reference implementation. Keep the same container paths after preparing a workspace. CSA uses shared Git objects from the read-only dataset; CodeQL databases are copied into the task workspace.

Dataset contents

• CSA: 159 tasks with source snapshots, captured build assets, reference implementations, and refinement environments. Shared runtime assets preserve restoration dependencies.
• CodeQL: four task bundles containing 30 Go, 30 Java, 30 JavaScript, and 51 Python tasks, with their captured analysis databases.
• scripts/prepare.py restores before/after scanning workspaces. The CSA bundle also includes scripts/rehydrate_benchmark_env.py for full refinement worktrees. Historical refinement assets contain absolute paths and symbolic links; see docs/environment.md for the restoration boundary.
• References and evaluator assets are maintainer material. During evaluation, expose only a prepared single-task workspace and its required runtime assets to the evaluated agent.

Integrity and validation

release-manifest.json records archive and part sizes, SHA-256 checksums, and the image fingerprint. environment/versions.json records the toolchain versions; the exported image is the direct restoration entry point. reports/validation.json records the checks actually performed and their limits.

Third-party source and tools retain their own license files. This repository does not assign a new blanket license to all bundled third-party content.

Paper

CheckerBench: Can Long-Horizon Agents Synthesize Static-Analysis Checkers?