Vulnerability Detection Datasets for Large Model Fine-tuning

It integrates multiple well-known benchmark datasets for code vulnerability detection, aiming to train models to identify security vulnerabilities in code through instruction fine-tuning. The dataset converts raw code snippets into a standard format of "instruction input output", suitable for supervised fine-tuning tasks, helping models learn code auditing and secure coding standards. The dataset consists of multiple subsets, each corresponding to a specific vulnerability detection benchmark or source, covering different types of vulnerabilities and code scenarios: BigVul: Contains real C/C++vulnerability code and its patched versions. CVEpixels: Based on the CVE database, it includes specific repair code for publicly available vulnerabilities. Devign: Focused on vulnerability detection in large-scale real-world projects. DiverseVul: emphasizes the diversity of vulnerability types. Juliet: The testing suite provided by NIST contains a large number of manually constructed vulnerability samples. MixVul: A publicly available dataset for software vulnerability detection, widely used in deep learning based vulnerability detection tasks ReVeal: A dataset built using Chromium and Debian repair commits The data files are stored in JSON List format.

The sample structure follows the standard instruction fine-tuning format and includes the following four core fields: instruction, input, output, and idx