DeepSource Benchmarks

Benchmark dataset evaluating code review and security analysis tools on the OpenSSF CVE Benchmark.

Benchmarked Tools

Last updated: February 8, 2026

• DeepSource
• Claude Code
• Codex
• CodeRabbit
• Cursor
• Devin
• Greptile
• Semgrep

Data Format

Judged Results (benchmarks/judged-results/)

Final evaluation results in JSONL format with fields:

• cve_id: CVE identifier
• variant: fixed or unfixed
• detected_issues: Issues found by the tool
• TP, FP, TN, FN: Classification metrics
• judge_reasoning: Explanation of the judgment

Processed Results (benchmarks/processed/)

Intermediate formatted results from each tool, normalized for comparison.

Raw Output (benchmarks/raw-output/)

Original tool outputs per CVE, preserving the exact response from each tool.

Archive

The archive/ directory contains prompts and data from earlier benchmark runs:

References

• OpenSSF CVE Benchmark