Zeezhu/grc-nist-csf2-qa-v1

Dataset Card for GRC NIST CSF 2.0 QA Dataset (v1)

This dataset contains question–answer and dialogue pairs based on the NIST Cybersecurity Framework (CSF) 2.0.
It is intended for training large language models to provide Governance, Risk, and Compliance (GRC) guidance.

---

Dataset Details

Dataset Sources

• Repository: https://huggingface.co/datasets/zeezhu/grc-nist-csf2-qa-v1
• Paper [optional]: NIST CSF 2.0 (official source: https://www.nist.gov/cyberframework)
• Demo [optional]: None yet

---

Uses

Direct Use

• Fine-tuning instruction-following models (e.g., LLaMA, Mistral, Qwen).
• Training chat assistants for cybersecurity and compliance.
• Educational tools for NGOs, SMEs, and consultants who need simplified NIST CSF guidance.

Out-of-Scope Use

• Not a replacement for legal, regulatory, or compliance advice.
• Not suitable as the only source for enterprise risk assessments.
• Should not be used for critical decision-making without human review.

---

Dataset Structure

• alpaca.jsonl: Instruction–input–output triples.

  • Fields:
    • instruction (string)
    • input (string, optional)
    • output (string)

• chatml.jsonl: Chat-formatted dialogues.

  • Fields:
    • messages: list of objects with role (system, user, assistant) and content (string).

• Splits: Current version contains a single combined dataset (~3,300 records). Users may create train/validation/test splits using random sampling or Unsloth utilities.

---

Dataset Creation

Curation Rationale

Created to provide structured Q&A data around the NIST CSF 2.0 framework, enabling AI models to learn compliance concepts in an approachable format.

Source Data

Data Collection and Processing

• Extracted from NIST CSF 2.0 JSON exports.
• Reformatted into Alpaca and ChatML formats for model fine-tuning.
• Quality checks: ensured valid JSONL (UTF-8), removed empty or duplicate entries.

Who are the source data producers?

• Primary source: NIST (official CSF 2.0 framework text).
• Reformatted and curated by the GRCora Project team.

Annotations

• No manual human annotations; transformation was automated.

Personal and Sensitive Information

• Dataset does not contain personal, sensitive, or private data.
• Content is limited to framework descriptions and compliance Q&A.

---

Bias, Risks, and Limitations

• Bias: Only covers NIST CSF 2.0; does not reflect other frameworks equally (e.g., ISO 27001, SOC 2).
• Risk: Users may mistake generated answers as authoritative compliance advice.
• Limitations: Simplified responses may not capture full legal/regulatory complexity.

Recommendations

• Use as training/evaluation data, not as a compliance checklist.
• Pair with domain expert review for production use.

---

Citation

BibTeX:

@misc{grc_nist_csf2_qa,
  title        = {GRC NIST CSF 2.0 QA Dataset (v1)},
  author       = {A.O.A},
  year         = {2025},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/zeezhu/grc-nist-csf2-qa-v1}
}
APA:
A.O.A. (2025). GRC NIST CSF 2.0 QA Dataset (v1) [Dataset]. Hugging Face. https://huggingface.co/datasets/zeezhu/grc-nist-csf2-qa-v1

Glossary

NIST CSF 2.0: National Institute of Standards and Technology Cybersecurity Framework, version 2.0.

GRC: Governance, Risk, and Compliance.

Alpaca format: Instruction–input–output JSONL structure used in fine-tuning.

ChatML: Chat-format schema for conversational AI training.

More Information

NIST CSF 2.0 official page: https://www.nist.gov/cyberframework

Dataset curated as part of GRCora (AI-driven GRC SaaS project).

Dataset Card Authors

A.O.A (curator, dataset creator)

Dataset Card Contact

Hugging Face: https://huggingface.co/zeezhu