fawazo/qwen2.5-coder-3b-pentest-gguf

Qwen2.5-Coder-3B Pentest - GGUF

GGUF quantizations of fawazo/qwen2.5-coder-3b-pentest optimized for Jetson Orin Nano (8GB).

Model Description

An AI pentesting assistant fine-tuned on 150K+ cybersecurity examples covering:

• OWASP Top 10 vulnerabilities
• MITRE ATT&CK framework
• API security testing
• Web application penetration testing

Output Format: JSON for automation

Quantizations

File	Size	RAM Needed	Recommended For
qwen2.5-coder-3b-pentest-q4_k_m.gguf	~1.8GB	~3GB	Jetson Orin Nano 8GB
qwen2.5-coder-3b-pentest-q5_k_m.gguf	~2.1GB	~4GB	Better quality
qwen2.5-coder-3b-pentest-q8_0.gguf	~3.4GB	~5GB	Best quality
qwen2.5-coder-3b-pentest-f16.gguf	~6GB	~8GB	Full precision

Usage on Jetson

With Ollama

# Download Q4_K_M (recommended for 8GB)
huggingface-cli download fawazo/qwen2.5-coder-3b-pentest-gguf qwen2.5-coder-3b-pentest-q4_k_m.gguf

# Create Modelfile
cat > Modelfile << 'EOF'
FROM ./qwen2.5-coder-3b-pentest-q4_k_m.gguf

SYSTEM """You are an expert penetration testing AI assistant. Analyze web traffic and respond with JSON:
{"action": "report|request|command|complete", ...}"""

PARAMETER temperature 0.3
PARAMETER num_ctx 2048
EOF

# Create and run
ollama create pentest-agent -f Modelfile
ollama run pentest-agent

With llama.cpp

./llama-cli -m qwen2.5-coder-3b-pentest-q4_k_m.gguf -ngl 99 -c 2048 -p "Analyze this request..."

Example Usage

Input:

Analyze this HTTP exchange:
REQUEST: GET /api/users?id=1
RESPONSE: {"user": "admin", "role": "administrator"}

Output:

{
  "action": "request",
  "method": "GET",
  "path": "/api/users?id=2",
  "reasoning": "Testing for IDOR - checking if user IDs are enumerable"
}

Training Details

• Base: Qwen/Qwen2.5-Coder-3B
• Method: SFT with LoRA (r=32)
• Dataset: 150K+ combined examples from Trendyol, Fenrir v2.0, pentest-agent
• Frameworks: OWASP, MITRE ATT&CK, NIST CSF

License

Apache 2.0 (inherits from base model and training datasets)