AlicanKiraz0's picture
Update README.md
531ea76 verified
metadata
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - cybersecurity
  - cve
  - vulnerability
size_categories:
  - 100K<n<1M

CVE Chat‑Style Multi‑Turn Cybersecurity Dataset (1999 – 2025)

1. Project Overview

This repository hosts the largest publicly available chat‑style, multi‑turn cybersecurity dataset to date, containing ≈ 300 000 Common Vulnerabilities and Exposures (CVE) records published between 1999 and 2025. Each record has been meticulously parsed, enriched, and converted into a conversational format that is ideal for training and evaluating AI and AI‑Agent systems focused on vulnerability analysis, threat intelligence, and cyber‑defense automation.

2. Key Highlights. Key Highlights

Feature Description
Records ~300 k CVE entries (1999‑2025)
Formats Covered CVE 4.0 (legacy) & CVE 5.0+ (modern)
Parsing Accuracy 100 % (validated)
Enrichments CVSS v2 & v3 metrics · CWE taxonomy · Affected‑product matrices · Expert system prompts
Conversation Depth Multi‑turn (System / User / Assistant)
Processing Pipeline Fully asynchronous, linearly scalable data‑engineering architecture
License Apache license 2.0

3. Intended Use Cases

  • Fine‑tuning LLMs for vulnerability triage and severity prediction.
  • Temporal trend analysis of vulnerability disclosures.
  • Retrieval‑Augmented Generation (RAG) and autonomous AI‑Agent pipelines.
  • Real‑time threat‑intelligence enrichment services.
  • Automated penetration‑testing (pentest) orchestration.

Benchmark Note
Early experiments with Llama 3.2 and Gemma models achieved 94 % accuracy on CVE class‑prediction tasks after full fine‑tuning on this dataset.

4. Dataset Structure

Each dialogue is stored as a single JSON Lines (.jsonl) object with three top‑level keys:

{
  "System": "You are a cybersecurity expert specializing in penetration testing, vulnerability research, and exploit development. Provide comprehensive technical analysis of CVE vulnerabilities with academic rigor and practical exploitation insights.",
  "User": "Provide a comprehensive technical analysis of CVE‑2010‑3763, including exploitation vectors, impact assessment, and remediation strategies.",
  "Assistant": "## CVE‑2010‑3763 Vulnerability Details

### CVE Metadata
- **CVE ID**: CVE‑2010‑3763
- **State**: PUBLISHED
..."
}

Field Reference

Key Type Description
System string System prompt that frames the assistant’s role and response style.
User string End‑user request or question.
Assistant string Model answer containing enriched CVE analysis and metadata.

Note: Multi‑turn conversations are represented as separate JSONL lines that share the same System context while User and Assistant evolve turn by turn.

5. Processing Pipeline. Processing Pipeline

  1. Source Aggregation – CVE XML feeds (4.0) + JSON feeds (5.0+).
  2. Asynchronous Parsing – Custom Rust & Python pipeline (Tokio + asyncio) for 100 % parsing success.
  3. Enrichment Layer – CVSS scoring, CWE classification, product‑matrix generation.
  4. Conversation Generation – Expert prompts injected to produce System / User / Assistant structure.
  5. Validation & QA – Schema checks, de‑duplication, manual spot‑checks.

6. Quick Start

Load with 🤗 datasets

from datasets import load_dataset

cve_chat = load_dataset("<username>/<repo_name>", split="train")
print(cve_chat[0])

Finetune Example (PEFT & QLoRA)

python train.py \
  --model "meta-llama/Meta-Llama-3-8B" \
  --dataset "<username>/<repo_name>" \
  --peft lora \
  --bits 4

7. Data Splits

Split Records Notes
train 240 000 80 %
validation 30 000 10 %
test 27 441 10 %

8. Contact

Contributions, feedback, and pull requests are warmly welcomed!