license: cc-by-nc-nd-4.0
Attack Vector Fine-tuning Dataset
This is a .json dataset of 3,660 synthetic instruction-tuning examples. It is intended to assist in training language models for automated red-teaming tasks, specifically focusing on the analysis of CVEs (Common Vulnerabilities and Exposures).
Source Data
The examples are derived from the official CVEListV5 repository, providing consensus vulnerability descriptions and standardized CVSS metrics.
Dataset Structure
The data uses valid CVE records to create three distinct types of training examples to enhance model capabilities:
- Attack Vector Identification: Teaches the model to deduce the primary attack vector (e.g., Network, Local, Physical) solely from the vulnerability description.
- Risk Assessment: Trains the model to estimate risk levels (Critical, High, Medium) and base scores based on the description.
- Executive Summary: Fine-tunes the ability to generate concise summaries combining the CVE ID, risk level, and attack vector.
Usage
The dataset is provided in a standard JSON format suitable for fine-tuning pipelines. It was generated with python to process raw CVE data, filter for quality (removing short descriptions or unknown vectors), and shuffle the entries for diverse training batches.