Datasets:

Modalities:
Text
Formats:
parquet
Languages:
Chinese
ArXiv:
License:
Vulnerability-CNVD / README.md
cedricbonhomme's picture
Update dataset card
fcfa115 verified
|
Raw
History Blame Contribute Delete
3.1 kB
metadata
language:
  - zh
license: cc-by-4.0
tags:
  - vulnerability
  - cybersecurity
  - cnvd
  - severity-classification
size_categories:
  - 100K-1M

Vulnerability-CNVD

Vulnerability descriptions and severity labels from the China National Vulnerability Database (CNVD), extracted via Vulnerability-Lookup.

Dataset structure

Field Type Description
id string CNVD identifier (e.g., CNVD-2025-03529)
title string Vulnerability title in Chinese
description string Vulnerability description in Chinese
severity string Severity level: 高 (High), 中 (Medium), or 低 (Low)
cve_id string Corresponding CVE identifier, if available (empty string if none)

Severity distribution

The dataset is imbalanced:

Severity Chinese Approximate share
High ~36%
Medium ~55%
Low ~9%

CVE overlap

Approximately 81% of CNVD entries have a corresponding CVE identifier. The overlap rate varies by year:

  • 2020-2021: 68-69% CVE mapping rate
  • 2022+: 91-97% CVE mapping rate

The ~19% of CNVD-only entries are concentrated in Chinese domestic software (PHP CMS, ERP systems). Western vendors (Adobe, Microsoft, IBM, Cisco) are largely absent from the CNVD-only subset.

Coverage and provenance

CNVD reserves 50,000-100,000 vulnerability IDs per year but publishes full details for only a fraction. The publication rate has declined significantly:

  • 2015: ~94% of reserved IDs have published details
  • 2023: ~4% of reserved IDs have published details

This decline coincides with China's Regulations on the Management of Security Vulnerabilities (RMSV), effective September 2021.

Entries without a description or severity label are excluded from this dataset.

Duplicate descriptions

CNVD reuses boilerplate descriptions across different vulnerability IDs (product-specific entries sharing the same text). When using this dataset for train/test splits, split on unique description text rather than on IDs to avoid data leakage. See VulnTrain#19 for details.

Source

Related models

References