# HackIDLE-NIST-Coder (GGUF)

A specialized cybersecurity LLM fine-tuned on 568 NIST publications, optimized for Ollama and llama.cpp.

## Model Details

**Base Model:** Qwen2.5-Coder-7B-Instruct
**Fine-tuning:** LoRA (11.5M parameters, 0.151% of base)
**Training Data:** 568 NIST cybersecurity documents (523,706 examples)
**Context Length:** 32,768 tokens
**License:** Apache 2.0

## Quantization Variants

| File | Size | Use Case | Perplexity |
|------|------|----------|------------|
| `hackidle-nist-coder-f16.gguf` | 14GB | Reference/source | Baseline |
| `hackidle-nist-coder-q8_0.gguf` | 7.5GB | Highest quality | ~0.1% loss |
| `hackidle-nist-coder-q5_k_m.gguf` | 5.1GB | High quality | ~0.5% loss |
| **`hackidle-nist-coder-q4_k_m.gguf`** | **4.4GB** | **Recommended** | ~1% loss |

## Usage

### With Ollama

Download and run:
```bash
ollama run ethanolivertroy/hackidle-nist-coder
```

Or create from this repo:
```bash
# Download GGUF
wget https://huggingface.co/ethanolivertroy/HackIDLE-NIST-Coder-GGUF/resolve/main/hackidle-nist-coder-q4_k_m.gguf

# Create Modelfile
cat > Modelfile << 'EOF'
FROM ./hackidle-nist-coder-q4_k_m.gguf

SYSTEM """You are HackIDLE-NIST-Coder, a cybersecurity expert with deep knowledge of NIST standards, frameworks, and best practices."""

PARAMETER temperature 0.7
PARAMETER num_ctx 32768
EOF

# Create model
ollama create hackidle-nist-coder -f Modelfile
```

### With llama.cpp

```bash
# Download GGUF
wget https://huggingface.co/ethanolivertroy/HackIDLE-NIST-Coder-GGUF/resolve/main/hackidle-nist-coder-q4_k_m.gguf

# Run inference
./llama-cli -m hackidle-nist-coder-q4_k_m.gguf \
    -p "What is Zero Trust Architecture according to NIST?" \
    -n 200 \
    --temp 0.7
```

### With LM Studio

1. Search for "hackidle-nist-coder" in LM Studio
2. Download Q4_K_M variant
3. Start chatting!

Or use the [MLX version](https://huggingface.co/ethanolivertroy/HackIDLE-NIST-Coder-MLX-4bit) for native Apple Silicon support.

## Expertise Areas

- NIST Cybersecurity Framework (CSF)
- Risk Management Framework (RMF)
- SP 800 series security controls (AC, AU, CA, CM, CP, IA, IR, MA, MP, PE, PL, PS, RA, SA, SC, SI, SR)
- FIPS cryptographic standards
- Zero Trust Architecture (SP 800-207)
- Cloud security (SP 800-210, SP 800-144)
- Supply chain risk management (SP 800-161)
- Privacy Framework

## Example Queries

```
"What is Zero Trust Architecture according to NIST SP 800-207?"
"Explain control AC-1 from NIST SP 800-53."
"What are the core components of the NIST Cybersecurity Framework?"
"How does NIST recommend implementing secure cloud architecture?"
"What is the Risk Management Framework process?"
```

## Training Details

**Dataset:** [`ethanolivertroy/nist-cybersecurity-training`](https://huggingface.co/datasets/ethanolivertroy/nist-cybersecurity-training)
- 523,706 training examples
- 568 source documents
- Smart chunking with sentence boundaries
- 5 extraction strategies: sections, controls, definitions, tables, semantic chunks

**Fine-tuning:**
- Method: LoRA with MLX (Apple Silicon)
- Training time: 3.5 hours on M4 Max
- Iterations: 1000
- Validation loss improvement: 45%
- Base model: Qwen2.5-Coder-7B-Instruct-4bit

## Performance

**Ollama (M4 Max, Q4_K_M):**
- Inference: 80-100 tokens/sec
- Memory: ~6GB
- Prompt processing: 50-100 tokens/sec

**llama.cpp (M4 Max, Q4_K_M):**
- Inference: 70-90 tokens/sec
- Memory: ~5GB

## Related Models

- **MLX Format:** [`ethanolivertroy/HackIDLE-NIST-Coder-MLX-4bit`](https://huggingface.co/ethanolivertroy/HackIDLE-NIST-Coder-MLX-4bit)
- **LM Studio:** [`ethanolivertroy/hackidle-nist-coder`](https://lmstudio.ai/ethanolivertroy/hackidle-nist-coder)
- **Ollama Library:** `ethanolivertroy/hackidle-nist-coder` (coming soon)

## Citation

If you use this model in your research or applications, please cite:

```bibtex
@software{hackidle_nist_coder,
  author = {Ethan Oliver Troy},
  title = {HackIDLE-NIST-Coder: A Fine-Tuned LLM for NIST Cybersecurity Standards},
  year = {2025},
  url = {https://huggingface.co/ethanolivertroy/HackIDLE-NIST-Coder-GGUF}
}
```

## License

This model is released under the Apache 2.0 license. NIST publications are in the public domain.

## Acknowledgments

- **NIST** for publishing comprehensive cybersecurity guidance
- **Qwen Team** for the exceptional Qwen2.5-Coder base model
- **llama.cpp** team for GGUF format and quantization
- **Ollama** for making local LLM deployment accessible