A compliance-specialized large language model fine-tuned by Nathan Maine for U.S. defense and regulated-industry cybersecurity compliance: CMMC 2.0, NIST SP 800-171 / 172 / 53, HIPAA, and DFARS.

Fine-tuned with QLoRA on an NVIDIA DGX Spark (GB10) over a curated, provenance-tracked compliance instruction dataset, and shipped as a quantized GGUF for fully local, air-gapped inference via Ollama or llama.cpp, with no cloud dependency.

Model details

Base model: Gemma 3 12B (Instruct)
Method: QLoRA (4-bit NF4)
Format: GGUF q5_k_m (~8.5 GB)
Domain: CMMC 2.0, NIST 800-171 / 172 / 53, HIPAA, DFARS
Training hardware: NVIDIA DGX Spark (GB10, 128 GB)

Usage (Ollama)

ollama run hf.co/Nathan-Maine/cmmc-expert-12b

Or download the GGUF and run it with llama.cpp.

Intended use and limitations

A research and drafting aid for compliance practitioners. It can be wrong; verify any control interpretation against the authoritative source documents before relying on it. Not legal advice.

License

This is a fine-tune of Google Gemma 3; use is subject to the Gemma Terms of Use.

Built by Nathan Maine. Training data and evaluation benchmarks: https://huggingface.co/Nathan-Maine . Pipeline and training code: https://github.com/NathanMaine/cmmc-compliance-ai-model