cmmc-expert-gemma4-31b-poc

CMMC Expert 31B (Gemma 4, Proof of Concept)

A proof-of-concept compliance-specialized large language model fine-tuned by Nathan Maine on Gemma 4 31B 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 4 31B
• Method: QLoRA (4-bit NF4)
• Format: GGUF q5_k_m (~21.9 GB)
• Status: Proof-of-concept variant in the CMMC-expert model line
• 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-gemma4-31b-poc

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

Intended use and limitations

A research and drafting aid for compliance practitioners, and a proof of concept for larger-model CMMC specialization. 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 4; 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