AMARA - 57 Real-World CVE Project Dataset for Code Security Auditing
Description
This dataset contains 57 real-world CVE projects used in the paper "AMARA: Adaptive Multi-Agent Routing for Project-Level Code Security Auditing". It spans three programming languages: 30 Java, 15 JavaScript/TypeScript, and 12 Python web applications, covering small (≤500 files), medium (500–2,000 files), and large (>2,000 files) scales. Each project is paired with a known CVE vulnerability and contains: (1) source_code/ — original source code at the vulnerable commit; (2) ground_truth/ — vulnerability location annotations (file path and line numbers) verified against CVE patches; (3) features/ — project-level static features (size, prescan risk, vulnerability type distribution) used for adaptive routing; (4) audit_results/ — audit outputs from all six collaboration modes for replication. The dataset supports reproducing AMARA's experiments on multi-agent code security auditing, including mode-level effectiveness, cost-effective
Files
Institutions
- Central China Normal UniversityHubei, Wuhan