VICBench
收藏资源简介:
VICBench是一个多语言代码漏洞检测基准数据集,由亚马逊云服务、匹兹堡大学和普渡大学联合构建。该数据集包含100个经过验证的漏洞引入提交(VICs),对应100个CVE,覆盖88个开源项目、3种编程语言(Python、Java、C++)和48种CWE类型,平均修复提交修改38.6行,VIC修改252.5行,显著大于先前工作。通过人类专家与VIC-Agent双重标注(κ=0.707)及专家验证,确保数据质量。该基准旨在评估漏洞检测工具的性能,填补现有数据集在语言多样性、补丁复杂性和项目范围上的空白,支持稳健的代码安全评估。
VICBench is a multilingual code vulnerability detection benchmark dataset jointly constructed by Amazon Web Services, University of Pittsburgh, and Purdue University. This dataset includes 100 verified vulnerability-introducing commits (VICs), corresponding to 100 Common Vulnerabilities and Exposures (CVE) entries, covering 88 open-source projects, three programming languages (Python, Java, C++), and 48 Common Weakness Enumeration (CWE) types. The average number of modified lines in repair commits is 38.6, while that in VICs is 252.5, which is significantly larger than those in prior research. Data quality is guaranteed via dual annotation by human experts and VIC-Agent (Cohen's κ=0.707) followed by expert validation. This benchmark is designed to evaluate the performance of vulnerability detection tools, fill the gaps of existing datasets in linguistic diversity, patch complexity, and project coverage, and support robust code security assessment.

- 1VICBench: A Multi-Language Benchmark for Code Vulnerability Detection亚马逊云服务; 匹兹堡大学; 普渡大学 · 2026年



