This is a training corpus for fine-tuning models — NOT an evaluation benchmark and NOT a compliance tool itself.
You will not train on the separate held-out benchmark questions (benchmark contamination); evaluate trained models against the separate v3 comprehensive benchmark instead.
Models trained on this data require qualified human expert review before any production use.
You will always review AI output before using it for compliance work and will not submit AI-generated compliance documentation without qualified human review.

Acknowledge the responsible-use terms below to access the training data.

EXPIRED VERSION. This release has been superseded by Nathan-Maine/cmmc-training-data-2026-09-30. Regulations change continuously — do not train compliance models on this version. It remains available for reproducibility and provenance only.

CMMC Training Data — 2026-09-16

A curated training corpus (train + validation splits) for fine-tuning small- and mid-size language models on CMMC 2.0, NIST SP 800-171/172, and related defense compliance frameworks. This is training data, not an evaluation benchmark — see the warning below on keeping training and evaluation strictly separate.

Version: 2026-09-16 Valid through: September 29, 2026 Next release: September 30, 2026 License: CC-BY-4.0 Author: Nathan Maine

Always-current link: Nathan-Maine/cmmc-training-data — cite and link the release index, which always points to the current release.

What This Is

A curated corpus of chat-formatted training examples covering CMMC 2.0, NIST SP 800-171, and related defense compliance frameworks. Examples are in OpenAI chat format (system/user/assistant) and are suitable for fine-tuning language models for compliance Q&A, definitional lookup, and framework summarization tasks.

It is intended so that researchers, compliance teams, and AI builders can:

Fine-tune their own domain-expert compliance models for research and evaluation
Benchmark LoRA/QLoRA pipelines against a known-good compliance corpus
See what production-quality compliance training data looks like
Build derivative datasets for adjacent regulatory domains

This corpus is the training member of a CMMC compliance-AI dataset family. The other members are held-out evaluation benchmarks (see Related Datasets). Keep the two strictly separated.

Content Overview

This release covers the following content areas:

Regulatory definitions — control family explanations, framework definitions, terminology, and standard-body reference material drawn from publicly available NIST, DoD, and CMMC publications
Framework summarization — condensed explanations of published standards, guidance documents, and program rules
Basic factual recall — Q&A over CMMC 2.0 levels, NIST SP 800-171 control structure, and DFARS clause references
Hallucination-resistance examples — examples that teach a model to refuse fabricated or non-existent regulatory constructs (e.g. "CMMC Level 4")

All content is derived from publicly available regulatory material. No CUI, no classified information, no customer data.

⚠️ Train Here, Evaluate Elsewhere — Avoid Benchmark Contamination

This is training data. Do not use it to measure model quality, and never train on benchmark questions.

Fine-tuning a model on a corpus and then evaluating that same model on overlapping questions inflates scores and hides real failure modes. To get a trustworthy read on a model trained from this corpus:

Train on this corpus (train + validation splits) only.
Evaluate against the separate held-out benchmark — use Nathan-Maine/cmmc-benchmark-v3-comprehensive-2026-q2, the authoritative comprehensive evaluation set.
Never add benchmark questions to your training mix. The benchmark sets are deliberately held out from this corpus; mixing them in defeats their purpose and produces unreliable scores. Every release is automatically screened against the benchmarks before publication.

The validation split here is for monitoring training (early stopping, loss tracking), not for final model evaluation. Final evaluation belongs to the held-out benchmark.

⚠️ AI Safety Disclaimer — Always Review Output

AI systems make mistakes. Always review AI-generated output before using it for any purpose.

This dataset is used to train AI systems that generate compliance guidance. Any AI system — including those trained on this data — can produce:

Factually incorrect information — even with high benchmark scores
Hallucinated citations — references to regulations, controls, or documents that do not exist
Outdated guidance — AI knowledge reflects training cutoff, not current regulations
Confident errors — AI often states wrong information with the same confidence as correct information
Plausible-sounding fabrications — responses that read like expert advice but are invented

Before using any AI output derived from this data for:

Compliance documentation (SSPs, POA&Ms, audit responses)
Regulatory submissions to DoD, NIST, or other agencies
Internal policy or procedure creation
Assessment preparation or C3PAO engagements
Legal or contractual decisions
Technical security implementation

You must:

Have a qualified human review every output — a compliance professional, security engineer, or subject matter expert
Verify citations independently — check that referenced controls, clauses, and publications exist and say what the AI claims
Cross-check against authoritative sources — NIST publications, DoD guidance, Federal Register, CMMC Assessment Guides
Document the review process — for audit purposes, maintain a record of who reviewed what and when
Never submit AI output directly — AI drafts are starting points for human work, not finished products

This is especially critical for CMMC and defense compliance because wrong answers can cause failed assessments, failed assessments can cost DoD contracts, and documentation enters the permanent record. C3PAO assessors verify human understanding, not AI output. The DoD holds contractors accountable for their submissions, not the tools they used.

Intended use: AI is a force multiplier for compliance professionals, not a replacement. Treat models trained on this data as drafting aids, not autonomous compliance systems. Every AI output is a draft for human review. The human stays accountable. The AI accelerates the work.

If you are using AI for compliance and do not have a qualified human in the review loop, stop. Either find one, or use a different tool.

⚠️ Version Expiration

Valid through: September 29, 2026 Next release: September 30, 2026

This training data is dated. CMMC regulations, DFARS clauses, and NIST publications update continuously. Training an AI on a frozen dataset produces a frozen model — one that may have been accurate at the time of training but grows stale the moment regulations change.

Updated every two weeks. Each release is rebuilt from a fresh scrape of authoritative sources (eCFR, NIST OSCAL catalogs, DoD publications, CISA KEV, FedRAMP) and incorporates:

New DFARS clauses and amendments
NIST SP 800-171/172 revisions and errata
CMMC Program Office guidance updates
Regulatory timeline updates (Phase rollouts, effective dates, etc.)
Assessment methodology changes

When a new release is published, this version is marked EXPIRED and remains available for reproducibility and provenance only. Always train on the current release — find it via the release index.

Dataset Details

Attribute	Value
Training split	7,822 examples
Validation split	869 examples
Total	8,691 examples
Format	JSON (OpenAI chat format: system/user/assistant)
Languages	English
License	CC-BY-4.0
Gating	Auto-approved (login + contact sharing required)

Frameworks Referenced

CMMC 2.0 — Levels 1, 2, 3
NIST SP 800-171 Rev 2 and Rev 3 — security requirements for CUI
NIST SP 800-172 — enhanced security requirements
DFARS 252.204-7012, 7019, 7020, 7021 — defense acquisition clauses
32 CFR Part 170 — CMMC Program Rule
NIST SP 800-53 Rev 5 — federal control catalog (referenced for cross-mapping)
Related: NIST CSF 2.0, HIPAA Security Rule, FIPS 140-3

Usage

from datasets import load_dataset

# Requires login to HuggingFace and auto-approved access
dataset = load_dataset("Nathan-Maine/cmmc-training-data-2026-09-16")

train = dataset["train"]          # 7,822 examples
validation = dataset["validation"] # 869 examples

# Example record structure
{
  "messages": [
    {"role": "system", "content": "You are a CMMC compliance expert..."},
    {"role": "user", "content": "What does AC.L2-3.1.1 require?"},
    {"role": "assistant", "content": "AC.L2-3.1.1 requires..."}
  ]
}

Intended Use

Good uses

Fine-tuning compliance AI models for research and internal evaluation
Validating LoRA/QLoRA training pipelines against a known-good corpus
Benchmarking fine-tuning hyperparameters on a domain-specific dataset
Building derivative datasets for adjacent compliance domains
Academic research on domain-specific LLM fine-tuning

Not intended for

Directly answering compliance questions (this is training data, not a model)
Evaluating model quality (use the separate held-out benchmark, never this corpus)
Producing submittable compliance documentation without human review
Replacing qualified C3PAO assessments
Classified or Sensitive Compartmented Information handling
Use as the sole source of compliance guidance

Related Datasets

Evaluation benchmarks (to score models trained on this corpus)

Nathan-Maine/cmmc-benchmark-v1-preview-2026-q2 — 46-question preview (methodology sampler, not for validation)
Nathan-Maine/cmmc-benchmark-v2-spotcheck-2026-q2 — 454-question spot check (triage only, not comprehensive)
Nathan-Maine/cmmc-benchmark-v3-comprehensive-2026-q2 — 1,273-question comprehensive evaluation (the authoritative standard)

These benchmark questions are held out from this training corpus. Evaluate trained models against them; do not train on them.

Citation

@dataset{maine-cmmc-training-data,
  author = {Maine, Nathan},
  title = {CMMC Training Data},
  year = {2026},
  url = {https://huggingface.co/datasets/Nathan-Maine/cmmc-training-data},
  note = {Biweekly release. This citation resolves to the release index, which links the current release (2026-09-16 at time of writing).}
}

Responsible AI Statement

This training corpus was curated for accuracy within CMMC and NIST compliance domains. However:

Regulations change continuously — a model trained on this data should be retrained on newer versions as releases are published
AI outputs trained from this data require human review before use in formal compliance documentation
The data reflects US defense contractor compliance requirements and may not apply to other jurisdictions
This dataset is not classified, not CUI, and contains only publicly available regulatory content

Changelog

2026-09-16 (Current)

Rebuilt from a fresh scrape of authoritative sources (snapshot 2026-09-15)
51 new examples added (46 train / 5 validation) after dedup and benchmark-contamination screening
Base corpus carried forward from Nathan-Maine/cmmc-training-data-2026-08-31
8691 total examples (7822 train / 869 validation)

Independent work by Nathan Maine. Training corpus only — evaluate trained models against the separate held-out benchmark, and never train on benchmark questions.

Downloads last month: 30

Total file size:

18.6 MB