AnamayaVyas/nvd-security-instructions

NVD Security Instructions

Plain English CVE analysis pairs for fine-tuning security-domain LLMs.

Dataset Summary

2,063 instruction-response pairs built from NVD CVE data (2023-2024). Each pair contains a raw CVE as input and a structured plain English analysis as output — designed to teach LLMs to explain vulnerabilities to developers with no security background.

Dataset Structure

Each record contains:

• text: Mistral instruction format string ready for fine-tuning

The instruction includes:

• CVE ID, description, CVSS score and vector
• Severity, published date, fix availability
• Affected products (CPE strings)
• CISA exploitation date if confirmed

The response includes:

• what_happened: plain English vulnerability explanation
• who_is_affected: specific software and versions
• how_bad_is_it: severity with attack vector context
• what_to_do: actionable remediation steps

How It Was Built

1. Fetched CVEs from NVD API (2023-2024, indices 220000-222000)
2. Oversampled CRITICAL and HIGH severity for balance
3. Applied quality filter (91% pass rate)
4. Enriched with CISA KEV exploitation dates
5. Generated outputs using Llama 3.1 8B via Ollama
6. Cleaned with Claude — removed 476 records with:
   • Hallucinated exploitation claims
   • Invented version numbers
   • Contradictory fix status
   • Raw CVSS string paste instead of plain English
7. Validated: 35/35 spot checks passed

Severity Distribution

Severity	Count	Percentage
CRITICAL	686	27.0%
HIGH	848	33.4%
MEDIUM	869	34.2%
LOW	134	5.3%

Known Limitations

• Outputs generated by Llama 3.1 8B — not human expert validated at scale
• Attack vector not always explicitly decoded in how_bad_is_it
• fix_available field may have false negatives
• Covers 2023-2024 CVEs only
• MITRE ATT&CK mapping not included (planned for v2)
• Security knowledge pairs (CVSS/EPSS/CWE) not included (planned for v2)

Intended Use

Fine-tuning LLMs for security domain CVE analysis tasks.

NOT intended for production security decision making without human expert review.

Citation

If you use this dataset please cite: Vyas, A. (2026). NVD Security Instructions Dataset. HuggingFace. https://huggingface.co/datasets/AnamayaVyas/nvd-security-instructions