TOPRA - 57 Real-World CVE Project Dataset for Code Security Auditing

Published: 2 November 2026| Version 1 | DOI: 10.17632/wyphrxv8jc.1
Contributor:

Description

This dataset supports the empirical evaluation in the paper "TOPRA: Pareto-Driven Topology Routing for Multi-Agent Code Auditing" (Journal of Systems and Software, preprint 2026). It contains 57 real-world open-source projects with confirmed CVE vulnerabilities and fine-grained ground-truth annotations, built for controlled benchmarking of multi-agent code-auditing systems, in particular studies of collaboration-topology design and adaptive routing strategies. Contents: - 57 projects across three languages: Java (30), JavaScript/TypeScript (15), Python (12). - 123 ground-truth vulnerability instances distributed over 110 distinct vulnerable source files. - Per-project annotations include CVE identifier, vulnerability type, CWE, file path, sink line range, entry chain, fix commit, and severity. - Project scale strata (by source-file count): Small (8), Medium (32), Large (17). - 68 distinct CVE entries referenced across the corpus. Structure: - projects/ — original source snapshots of the 57 CVE projects, organized by language (java, javascript, python), one <project>.zip each. - ground_truth/ — fine-grained annotations, one ground_truth_<project>.json per project, organized by language. - manifest.csv — canonical index (one row per project) with paths to source archive and ground-truth file. - dataset_metadata.csv / metadata.json — per-project metadata summary. The dataset is released for academic and defensive security research only (vulnerability detection, code auditing, benchmarking of multi-agent auditing systems). License: CC BY 4.0. When using this dataset, please cite the accompanying paper and dataset DOI (https://doi.org/10.17632/wyphrxv8jc.1). The repository is split into two folders: - `ground_truth`: Vulnerability annotations, metadata, and statistical records for each project. - `projects`: Original source code of the collected CVE projects. This dataset is constructed to support the evaluation of multi-agent collaboration topologies and adaptive routing strategies for vulnerability detection. It is used in the paper TOPRA: Pareto Routing for Multi-Agent Code Auditing, enabling controlled, multi-objective benchmarking of LLM-based code auditing agents.

Files

Categories

Computer Science, Artificial Intelligence, Cybersecurity

Licence