ivitopow/SecuCoder-GGUF

SecuCoder — GGUF

Quantized GGUF version of SecuCoder, a fine-tuned Llama 3.1 8B Instruct model for secure Python code generation and vulnerability remediation.

For full model details, training methodology, and evaluation results, see the main model card.

---

Available Files

File	Quantization	Size	Use case
secucoder-Q4_K_M.gguf	Q4_K_M	~4.6 GB	Recommended — best balance of quality and size

---

Usage with Ollama

1. Download the Modelfile from this repo and create the model:

ollama create secucoder -f Modelfile

2. Run it:

ollama run secucoder

3. Or via API:

curl http://localhost:11434/api/generate -d '{
  "model": "secucoder",
  "prompt": "Fix the security vulnerability in this Python code.\n\n```python\nname = request.args.get(\"name\")\nresp = make_response(\"Your name is \" + name)\n```\n\nCWE: CWE-079",
  "stream": false
}'

---

Usage with llama.cpp

./llama-cli \
  -m secucoder-Q4_K_M.gguf \
  --ctx-size 4096 \
  --temp 0.1 \
  --top-p 0.9 \
  -p "You are a secure Python assistant. Fix the vulnerability in this code: ..."

---

Recommended Parameters

Parameter	Value
temperature	0.1
top_p	0.9
num_ctx	4096
num_predict	3072

---

System Prompt

You are a secure Python assistant. Help identify, explain, and fix security issues in Python code. Prefer safe, practical, and production-ready solutions.

---

Evaluation

The full SecuCoder system (Q4 + structured prompting + RAG) achieves an overall score of 77.11 vs 60.34 for the untuned Llama 3.1 8B baseline — a +27.8% improvement measured by weighted static analysis findings (Bandit + Semgrep).

Variant	Overall Score
Llama 3.1 8B Instruct (baseline)	60.34
SecuCoder Q4 (this model)	61.46
SecuCoder Q4 + structured prompt	64.46
SecuCoder Q4 + structured prompt + RAG	77.11

---

Related

Resource	Link
Full model (safetensors)	ivitopow/secucoder
Training dataset	ivitopow/secucoder
Base model	meta-llama/Llama-3.1-8B-Instruct

---

License

Released under CC BY-NC-SA 4.0. Built on Llama 3.1, subject to Meta's Llama 3 Community License.