CyberArchitect-Qwen-2.5-7B-GGUF is a specialized small language model fine-tuned for cybersecurity architecture review, infrastructure-as-code analysis, and threat modeling.
The model is trained to reason step-by-step before producing conclusions, reducing false positives commonly seen in general-purpose LLMs.

Author: Syamchand
Base Model: Qwen/Qwen2.5-7B-Instruct
Format: GGUF (Q4_K_M)
Fine-tuning Method: LoRA (Unsloth)
Intended Runtime: CPU / Mac / consumer hardware

Intended Use

This model is intended for defensive security (Blue Team) and educational or research purposes, including:

Kubernetes & cloud configuration reviews
Infrastructure-as-Code (IaC) security analysis
Architectural threat modeling
Secure coding logic validation
Identifying misconfigurations and insecure design patterns

Not Intended For

Automated vulnerability exploitation
Offensive security or malware development
Replacing professional security audits or production SAST tools

Model Behavior

The model is trained to follow a structured reasoning pattern, explicitly analyzing:

Data flow and trust boundaries
Authentication and authorization logic
Configuration context (cloud, IAM, K8s)
Security best practices vs actual implementation

It prioritizes context-aware reasoning over keyword matching.

Evaluation

Evaluation Setup

The model was evaluated against the base Qwen 2.5 7B Instruct using a curated internal dataset containing:

Secure code samples
Intentionally vulnerable code
Kubernetes YAML & cloud configuration files

Metrics (Qualitative)

Metric	Base Model	CyberArchitect
False Positive Rate	High	Low
Secure Pattern Recognition	Inconsistent	Strong
Reasoning Coherence	Generic	Structured & contextual
IaC / Cloud Understanding	Limited	Strong

Key Findings

Significantly reduced false positives on secure SQL and JWT handling
Improved reasoning on Kubernetes Secrets vs ConfigMaps
Better identification of architectural risks rather than surface-level issues

Limitations

Optimized primarily for infrastructure and architecture, not general conversation
May over-analyze authentication logic in small scripts
Reasoning quality depends on correct system prompt usage

Usage

System Prompt (Recommended)

You are a Cyber Architect. You must always start your response with a
'Thinking Process:' section where you analyze the request step-by-step,
followed by your final 'Analysis:'.

Example (llama-cpp-python)

python

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="Syamchand/CyberArchitect-Qwen-2.5-7B-GGUF",
    filename="Qwen2.5-7B-Instruct.Q4_K_M.gguf",
    n_ctx=4096,
    n_threads=4
)

Training Details

Fine-tuning: LoRA via Unsloth
Dataset: Custom Chain-of-Thought cybersecurity reasoning dataset
Hardware: NVIDIA T4 GPU
Quantization: GGUF via llama.cpp

Ethical Considerations

This model is designed strictly for defensive cybersecurity analysis.
It should not be used to generate exploits, bypass systems, or conduct malicious activities.

License

Apache 2.0