What It Does#
A fine-tuned LLM specialized in NIST cybersecurity guidance. Ask it about SP 800-53 controls, CSF 2.0 functions, Zero Trust architecture, FIPS cryptographic requirements, or any of the 596 NIST publications it was trained on.
Available on Ollama.
Quick Start#
ollama run etgohome/hackidle-nist-coder:v1.1 "What is Zero Trust Architecture?"That’s it. Pulls the model and starts answering.
Training Details#
| Detail | Value |
|---|---|
| Base Model | Qwen2.5-Coder-7B-Instruct (4-bit) |
| Method | LoRA (11.5M trainable params, 0.151% of total) |
| Training Data | 530,912 examples from 596 NIST documents |
| Hardware | Apple M4 Max, 128GB RAM |
| Training Time | ~3.5 hours (1,000 iterations) |
| Inference Speed | 130-160 tok/s on M4 Max |
| Context Window | 32K tokens |
| License | CC0 1.0 (public domain) |
What It Covers#
- SP 800-53 Rev. 5 - Full security and privacy control catalog
- Cybersecurity Framework 2.0 - Including the new GOVERN function
- Zero Trust Architecture (SP 800-207)
- Risk Management Framework
- FIPS cryptographic standards
- Post-quantum cryptography migration guidance
- Privacy Framework v1.0
- Cloud security guidance
- IoT cybersecurity and supply chain risk management
Model Variants#
| Format | Size | Use Case |
|---|---|---|
| Ollama (Q4_K_M) | 4.7 GB | Quickest setup, ollama run |
| GGUF Q4_K_M | 4.4 GB | llama.cpp, LM Studio |
| GGUF Q5_K_M | 5.1 GB | Better quality, slightly larger |
| GGUF Q8_0 | 7.5 GB | High quality inference |
| GGUF F16 | 14 GB | Full precision |
| MLX (4-bit) | — | Apple Silicon optimized |
Related#
- Open Source Security Compliance AI & ML - The broader collection of datasets and models this is part of
