A document classification benchmark dataset on cybersecurity policy domain

Document classification in the cybersecurity domain is essential for the efficient identification of emerging threats, technologies, and policy trends. However, existing datasets and language models are biased toward technical, attack-oriented texts such as CTI reports and CVE descriptions, and there is a critical lack of data for the cybersecurity policy domain. To address this gap, we introduce the first benchmark dataset for cybersecurity policy document classification, comprising 1,723 expert-verified samples categorized into three labels: Strategy (490), Legislation (415), and Guideline (818). The dataset was constructed through automated web crawling from five sources: UNIDIR, CIPedia, NIST, ENISA, and ACSC. To demonstrate the dataset’s effectiveness, we conducted classification experiments using three BERT-based models. With sliding window-based data augmentation applied, XLM-RoBERTa-base achieved the highest document accuracy of 99.99% and an F1 score of 99.36%. The dataset and code are publicly available to facilitate future research.