jason-oneal/mitre-stix-cve-exploitdb-dataset-alpaca-chatml-harmony

MITRE+NVD+ExploitDB Dataset (Alpaca/ChatML/Harmony)

A dataset for training AI assistants/agents on vulnerability analysis and pentesting Q&A. It is built by the pentestds pipeline, which fetches and merges data from MITRE CVE, NVD (CVSS enrichment), ExploitDB, and a small set of HuggingFace datasets. Provenance is recorded for every entry, and the pipeline emits Alpaca, ChatML, and Harmony JSONL files.

Dataset Summary

This dataset is designed for training AI agents to understand and perform penetration testing tasks. It is built by an automated Python pipeline that:

• Downloads CVE data from MITRE
• Streams CVEs from NVD and extracts CVSS where available
• Indexes ExploitDB and links exploits to CVEs
• Loads and merges additional datasets from HuggingFace (e.g., MITRE ATT&CK reasoning, TTP mapping, SecureCode v2)
• Validates records using Pydantic schemas for Alpaca and ChatML formats
• Tracks provenance for every record
• Outputs Alpaca, ChatML, and Harmony JSONL files
• Uploads the dataset to HuggingFace when TOKEN is configured

Included data types:

• CVE Data: Real vulnerability information from MITRE
• CVSS: NVD CVSS metrics (when present)
• Exploit Code: Proof-of-concept exploit references/snippets from ExploitDB
• Secure Coding: Multi-turn secure coding dialogues and vulnerability remediations
• Red Team Techniques: MITRE ATT&CK-aligned reasoning data
• Security Mappings: TTP mapping scenarios

Supported Tasks

• Vulnerability Analysis: Understanding and explaining CVEs
• Exploit Development: Writing and understanding exploit code
• Pentesting Methodology: Planning and executing penetration tests
• Red Team Operations: Advanced persistent threat simulation
• Tool Usage: Understanding cybersecurity tools and commands

Dataset Structure & Pipeline

The dataset is available in multiple formats (Alpaca, ChatML, Harmony), generated by the pipeline. The build process can be triggered by running pentestds build.

Naming conventions

Repository name

By default, the builder uploads to:

• jason-oneal/mitre-stix-cve-exploitdb-dataset-alpaca-chatml-harmony

This name comes from the builder’s USERNAME + a fixed suffix.

File names

Each format is written as JSONL and split into train/validate:

• Alpaca: alpaca_train.jsonl, alpaca_validate.jsonl
• ChatML: chatml_train.jsonl, chatml_validate.jsonl
• Harmony: harmony_train.jsonl, harmony_validate.jsonl (each line is {"text": "...raw Harmony tokens..."})

Metadata files:

• Dataset card: README.md (uploaded from the builder repo’s card.md)
• Citation: CITATION.cff
• Provenance: provenance.json

Alpaca Format

{
    "instruction": "Explain CVE-2023-1234",
    "input": "",
    "output": "CVE-2023-1234 is a critical vulnerability in Example Software..."
}

ChatML Format

{
    "messages": [
        {"role": "user", "content": "Explain CVE-2023-1234"},
        {"role": "assistant", "content": "CVE-2023-1234 is a critical vulnerability in Example Software..."}
    ]
}

Data Sources

Source	Type	License	Records	URL
MITRE CVE Database	CVE	MITRE CVE License	varies	https://cve.mitre.org/
National Vulnerability Database	CVE	NIST License	varies	https://nvd.nist.gov/
Exploit Database	EXPLOIT	ExploitDB License	varies	https://www.exploit-db.com/
MITRE ATT&CK Reasoning	REDTEAM	Apache-2.0	~300	https://huggingface.co/datasets/cobo512/Mitre-ATTACK-reasoning-dataset
Security TTP Mapping	SCENARIO	Apache-2.0	~500	https://huggingface.co/datasets/tumeteor/Security-TTP-Mapping
SecureCode v2	SCENARIO	Apache-2.0	~430	https://huggingface.co/datasets/scthornton/securecode-v2

Data Processing

Content Validation

All data undergoes content validation to ensure quality and consistency.

Content Cleaning

Content is cleaned and validated to ensure proper formatting and length.

Validation

All records are validated against Pydantic schemas to ensure data quality and format consistency.

Train/Validation Split

The dataset is split using deterministic hash-based partitioning with optional stratification by source or license type.

Usage

Loading the Dataset

from datasets import load_dataset

repo_id = "jason-oneal/mitre-stix-cve-exploitdb-dataset-alpaca-chatml-harmony"

# Load specific format files directly
alpaca_train = load_dataset(repo_id, data_files={"train": "alpaca_train.jsonl"})["train"]
chatml_train = load_dataset(repo_id, data_files={"train": "chatml_train.jsonl"})["train"]
harmony_train = load_dataset(repo_id, data_files={"train": "harmony_train.jsonl"})["train"]

Example Usage

# Get a sample record from alpaca format
alpaca_sample = alpaca_train[0]
print(f"Instruction: {alpaca_sample['instruction']}")
print(f"Output: {alpaca_sample['output']}")

# Get a sample record from chatml format
chatml_sample = chatml_train[0]
for message in chatml_sample['messages']:
    print(f"{message['role']}: {message['content']}")

Training Example

from transformers import AutoTokenizer, AutoModelForCausalLM

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained("microsoft/DialoGPT-medium")
tokenizer = AutoTokenizer.from_pretrained("microsoft/DialoGPT-medium")

# Prepare data
def format_prompt(example):
    if 'messages' in example:
        # ChatML format
        return "\n".join([f"{msg['role']}: {msg['content']}" for msg in example['messages']])
    else:
        # Alpaca format
        return f"Instruction: {example['instruction']}\nOutput: {example['output']}"

# Tokenize and train
# ... training code ...

Data Quality

Validation Results

• Total Records Processed: ~250,000
• Valid Records: ~245,000 (98%)
• Duplicates Removed: ~5,000
• Content Cleaned: ~1,000

Quality Metrics

• Schema Compliance: 100% (all records pass Pydantic validation)
• Total Records: ~3,730+ (combined from all sources)
• Source Attribution: 100% (all records have provenance tracking)

Limitations and Biases

Known Limitations

1. Language: Dataset is primarily in English
2. Temporal Coverage: CVE data limited to available years
3. Tool Coverage: Focus on common pentesting tools
4. Scenario Diversity: Limited to available pentesting scenarios

Potential Biases

1. Source Bias: Heavy reliance on MITRE/NVD for vulnerability data
2. Tool Bias: Focus on popular open-source tools
3. Geographic Bias: Primarily Western cybersecurity practices

Citation

If you use this dataset in your research, please cite:

@misc{mitre_nvd_exploitdb_dataset,
  title={MITRE+NVD+ExploitDB Dataset (Alpaca/ChatML/Harmony)},
  author={Jason O'Neal},
  year={2024},
  url={https://huggingface.co/datasets/jason-oneal/mitre-stix-cve-exploitdb-dataset-alpaca-chatml-harmony}
}

License

This dataset is licensed under Apache-2.0. Individual data sources retain their original licenses:

• MITRE CVE: Public domain
• ExploitDB: Various licenses per exploit
• HuggingFace Datasets: Apache-2.0

Contributing

Contributions are welcome! Please see the repository for contribution guidelines.

Updates

This dataset is updated daily via automated GitHub Actions workflows. Each update includes:

• Latest CVE data from MITRE
• Latest CVSS enrichment from NVD (when available)
• New exploits from ExploitDB
• Updated secure coding scenarios from SecureCode v2
• Updated MITRE ATT&CK reasoning and TTP mapping datasets (when accessible)

Contact

For questions or issues:

• GitHub Issues: Repository Issues
• Email: jason.allen.oneal@gmail.com

---

Generated by the Pentest Dataset Builder Pipeline