pretty_name: Cyber Security Evidence Dataset — CISA KEV Structured CC0 Layer
language:
- en
license: cc0-1.0
tags:
- cybersecurity
- vulnerability-management
- cve
- cisa-kev
- defensive-security
task_categories:
- text-classification
size_categories:
- 1K<n<10K
Cyber Security Evidence Dataset — CISA KEV Structured CC0 Layer
This configuration is the structured CISA Known Exploited Vulnerabilities (KEV) layer of the broader Cyber Security Evidence Dataset project. It contains 1,687 deterministic records generated from the official CISA KEV database snapshot.
What is included
The records contain the official KEV database fields: CVE identifier, vendor/project, product, vulnerability name, short description, dates, known-ransomware-use flag, required action, notes, CWE values where present, source provenance, hashes, and project-generated quality metadata. The canonical published Parquet files provide a flattened representation of the structured records.
The text field is composed only from the official KEV CVE identifier and short description. The dataset contains no vendor advisory pages, screenshots, PDFs, exploit code, linked web-page text, or cached third-party content. Empty entity, relation, and ATT&CK-link arrays are intentional: this configuration is a structured vulnerability catalog layer, not a human-adjudicated evidence-extraction corpus.
License and attribution
The official KEV database is distributed under the CISA source-specific CC0 1.0 license. This release preserves that notice and attribution. Information at third-party links included in KEV records remains governed by the terms of those third parties. This repository does not use CISA or DHS logos or seals and does not imply endorsement by CISA or DHS.
The source URLs, exact license URL, record counts, split counts, and published-file format are recorded in manifest.json and NOTICE.md. Each record also carries its available provenance fields. The project-authored transformations are limited to deterministic field mapping, serialization, and metadata.
Splits and reproducibility
The release provides chronological train, validation, and test splits with 1,181, 253, and 253 records, respectively. Parquet is the canonical published format for all three splits. Each split was generated from the same source snapshot and passed schema, duplicate, language, content-policy, and loader checks.
Limitations
This layer does not contain human-reviewed entity extraction, evidence spans, relation annotations, or MITRE ATT&CK mappings. It does not represent the private CISA ICS or vendor-advisory layers. The presence of a CVE in KEV indicates inclusion in the official catalog; it is not an independent claim by this project about exploitation details beyond the source field.
Safety
The release is intended for defensive vulnerability management, catalog modeling, retrieval, and provenance-aware data engineering. It intentionally excludes operational exploit instructions and linked third-party content.