ethanolivertroy/cmmc-training-balanced

CMMC Training Dataset - Balanced Variant

Dataset Description

This is the Balanced variant of the CMMC (Cybersecurity Maturity Model Certification) training dataset, containing 2,790 high-quality training examples with balanced coverage across all 17 CMMC domains.

Dataset Characteristics

• Total Examples: 2,790 (2,232 train / 558 validation)
• Source Documents: 71 NIST publications
• CMMC Levels Covered: Level 1, Level 2, Level 3
• CMMC Domains: All 17 domains (evenly distributed)
• Format: JSONL with chat-formatted messages
• Embeddings: 1536-dimensional vectors (OpenAI text-embedding-3-small)
• License: Public Domain (NIST documents are US Government works)

What Makes This "Balanced"?

The Balanced variant provides equal representation across all 17 CMMC domains, ensuring comprehensive coverage without bias toward any particular security area.

Domain Distribution (Perfectly Balanced)

Each of the 17 CMMC domains has approximately 840 examples:

• Access Control (AC): 840 examples
• Audit and Accountability (AU): 840 examples
• Awareness and Training (AT): 840 examples
• Configuration Management (CM): 823 examples
• Identification and Authentication (IA): 840 examples
• Incident Response (IR): 840 examples
• Maintenance (MA): 840 examples
• Media Protection (MP): 840 examples
• Personnel Security (PS): 840 examples
• Physical Protection (PE): 840 examples
• Planning (PL): 840 examples
• Risk Assessment (RA): 840 examples
• Security Assessment (CA): 840 examples
• Supply Chain Risk Management (SR): 825 examples
• System and Communications Protection (SC): 840 examples
• System and Information Integrity (SI): 840 examples
• System and Services Acquisition (SA): 840 examples

Note: Domain counts represent the number of examples tagged with each domain. Since examples can be tagged with multiple domains, the sum of domain counts exceeds the total number of examples (2,790).

This balanced distribution ensures your model learns all CMMC domains equally well.

CMMC Level Distribution

All Levels:         2,247 examples (80.5%)
Level 3 (Advanced): 301 examples (10.8%)
Level 2 (Advanced): 124 examples (4.4%)
Level 1 (Foundational): 118 examples (4.2%)

Source Documents (71 total)

The Balanced variant includes:

Core Foundation (14 documents):

• NIST SP 800-171 R3 (CMMC Level 2)
• NIST SP 800-172 R3 (CMMC Level 3)
• NIST SP 800-53 R5 (Master controls)
• Assessment procedures and supplementary guidance

Domain-Specific Publications (57 additional documents):

• Selected from 596 NIST publications using domain keyword matching
• Covers emerging topics: cloud security, IoT, supply chain, incident response
• Includes practical implementation guides and case studies

Examples:

• SP 800-61 (Incident Response)
• SP 800-92 (Audit Log Management)
• SP 800-115 (Security Testing)
• SP 800-137 (Continuous Monitoring)
• SP 800-161 (Supply Chain Risk Management)

Dataset Structure

JSONL Training Files

Each example follows the chat format:

{
  "messages": [
    {
      "role": "system",
      "content": "You are a cybersecurity expert specializing in CMMC..."
    },
    {
      "role": "user",
      "content": "How should incident response be implemented for CMMC Level 2?"
    },
    {
      "role": "assistant",
      "content": "According to NIST SP 800-61 R2, incident response for CMMC Level 2..."
    }
  ],
  "metadata": {
    "source": "NIST SP 800-61 R2",
    "cmmc_level": "2",
    "cmmc_domain": "Incident Response",
    "type": "domain_specific"
  }
}

Vector Embeddings

Pre-computed embeddings using OpenAI's text-embedding-3-small model:

• Format: Parquet files with 1536-dimensional vectors
• Files: embeddings_train.parquet, embeddings_valid.parquet
• Size: 34.5 MB total (27.5 MB train + 7.0 MB validation)
• Cost: $0.01 (674,102 tokens processed)

FAISS Indexes

Ready-to-use vector similarity search indexes:

• L2 distance indexes: faiss_train_l2.index, faiss_valid_l2.index
• Cosine similarity indexes: faiss_train_cosine.index, faiss_valid_cosine.index

Q&A Generation Strategies

Examples were generated using 5 complementary strategies:

1. Section-based Q&A: Questions from document sections
2. Control-based Q&A: NIST control requirements (3.1.1 format)
3. CMMC-specific Q&A: Level-focused questions (L1/L2/L3)
4. Domain-specific Q&A: Questions per CMMC domain (balanced sampling)
5. Semantic chunking: General content with context preservation

Use Cases

The Balanced dataset is ideal for:

• Comprehensive CMMC training: Equal coverage of all 17 domains
• Domain-agnostic models: No bias toward specific security areas
• Full compliance coverage: Suitable for general CMMC consulting
• RAG systems: Balanced retrieval across all domains
• Security assessment tools: Complete domain coverage
• Training multiple specialists: Each domain well-represented

Dataset Statistics

Source Documents:         71
Total Examples:           2,790
Training Examples:        2,232 (80%)
Validation Examples:      558 (20%)
Avg Example Length:       ~242 tokens
Total Tokens Embedded:    674,102
Embedding Cost:           $0.01 USD
Domain Balance:           99.8% (within 2% variance)

Advantages Over Other Variants

vs. Core (14 docs, 1.2K examples):

• 2.2x more examples
• 5x more source documents
• Better domain balance (Core is weighted toward SP 800-171/172)
• More diverse use cases and scenarios

vs. Comprehensive (381 docs, 11.3K examples):

• Faster training (75% fewer examples)
• Better signal-to-noise ratio (curated selection)
• Equal domain representation (Comprehensive may be imbalanced)
• Lower computational cost

Quick Start

Load JSONL Data

import json

# Load training data
with open('train.jsonl', 'r') as f:
    train_data = [json.loads(line) for line in f]

# Example: Check domain distribution
from collections import Counter
domains = [ex['metadata'].get('cmmc_domain', 'Unknown')
           for ex in train_data if 'metadata' in ex]
print(Counter(domains))

Load Embeddings

import pandas as pd
import numpy as np

# Load embeddings
df = pd.read_parquet('embeddings_train.parquet')

# Access embeddings as numpy array
embeddings = np.vstack(df['embedding'].values)
texts = df['text'].tolist()

print(f"Embeddings shape: {embeddings.shape}")  # (2232, 1536)

Use FAISS Index

import faiss

# Load FAISS index
index = faiss.read_index('faiss_train_cosine.index')

# Search for similar content
query_embedding = ... # your query vector (1536-dim)
k = 5  # number of results
distances, indices = index.search(query_embedding.reshape(1, -1), k)

# Get similar texts
for i, idx in enumerate(indices[0]):
    print(f"{i+1}. {texts[idx][:100]}...")

Related Datasets

This is part of a family of 3 CMMC datasets:

• Core: 14 docs, 1.2K examples - Essential CMMC foundation
• Balanced (this dataset): 71 docs, 2.8K examples - Domain-balanced coverage
• Comprehensive: 381 docs, 11.3K examples - Complete NIST CMMC library

When to Use Balanced vs. Others

Choose Balanced if:

• You need equal representation across all 17 CMMC domains
• You want comprehensive coverage without excessive examples
• You're building a general-purpose CMMC assistant
• You want faster training than Comprehensive
• You need more diversity than Core

Choose Core if:

• You only care about SP 800-171/172 fundamentals
• You want the fastest training possible
• You're focused on core CMMC requirements only

Choose Comprehensive if:

• You need maximum context and coverage
• You're building an exhaustive knowledge base
• Training time/cost is not a constraint
• You want every NIST CMMC-related publication

Citation

If you use this dataset, please cite:

@dataset{cmmc_balanced_2025,
  title={CMMC Training Dataset - Balanced Variant},
  author={Troy, Ethan Oliver},
  year={2025},
  publisher={HuggingFace},
  note={Derived from NIST Special Publications (Public Domain)}
}

License

Public Domain - This dataset is derived from NIST Special Publications, which are works of the US Government and not subject to copyright protection in the United States.

Acknowledgments

This dataset is built from publications by the National Institute of Standards and Technology (NIST), Computer Security Resource Center.

Dataset Version

• Version: 1.0
• Created: 2025
• Source: NIST CSRC Publications
• Processing: Docling + custom CMMC-aware data preparation
• Balancing: Domain-based keyword matching with weighted sampling

Contact

For questions or issues, please open an issue on the GitHub repository.