Vulnerability Scores: 709k Real-World Software Vulnerabilities for Severity Prediction

709,321 real-world software vulnerabilities train VLAI, a transformer model for automated severity classification. The dataset is presented in the paper 'VLAI: A RoBERTa-Based Model for Automated Vulnerability Severity Classification'. It was created by CIRCL and last updated on Hugging Face in May 2026.

Use Cases

• Training transformer models for vulnerability severity prediction based on text descriptions.
• Benchmarking automated triage systems against a large corpus of real-world Common Vulnerabilities and Exposures (CVE) entries.
• Analyzing patterns in vulnerability descriptions to improve consistency in severity labeling.

Strengths

• Contains 709,321 real-world vulnerability records, providing substantial scale for model training.
• Explicitly designed for training and evaluating VLAI, a transformer-based model for a specific NLP task.
• Sourced from the CVE Program, indicating a connection to a canonical vulnerability database.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Freshness should be verified as the last update timestamp is from the future (2026-05-28).

Provenance

Source

CIRCL

Collection Method

Likely aggregated and enriched from the CVE Program (cvelistv5).

Freshness

Last updated 2026-05-28 06:12:19