This is a custom cybersecurity fine-tune of Qwen2.5-3B-Instruct, optimized for local hardware and made using Unsloth.
It is designed to act as a Tier-3 SOC Analyst, strictly enforcing a Phase 1 (Core Analysis) and Phase 2 (Tactical Mitigation) framework when analyzing threat logs.
🔗 GitHub Repository: srihari2020/cybermind
| Parameter | Value |
|---|---|
| framework | Unsloth |
| rank (r) | 16 |
| lora_alpha | 32 |
| quantization | 4-bit (NF4) |
| dataset_size | 2,485 high-quality instruction pairs |
| context_window | 8192 |
Special training note: The training path for this release ran pure supervised fine-tuning (SFT) using Unsloth. The teacher data was derived from CISA KEV and MITRE ATT&CK frameworks. No coding data was included; the emphasis was purely on structured derivation, threat detection, and reliable final-answer construction formatted for Security Operations Centers.
Strong log analysis performance and the best average formatting adherence among tested baselines on NVIDIA RTX 5060.
| FORMAT ADHERENCE | AVERAGE LATENCY | GAIN VS BASE (QWEN2.5) |
|---|---|---|
| 100.0% | 2.04s | +60.4% speed |
| Zero Hallucinations | per log on RTX 5060 | -3.12s latency drop |
You can run this model locally with complete privacy using Ollama and the provided Modelfile.
git lfs install
git clone [https://huggingface.co/Srihari2020/CyberMind-4B-GGUF](https://huggingface.co/Srihari2020/CyberMind-4B-GGUF)
ollama create cybermind-4b -f CyberMind-4B-GGUF/Modelfile
ollama run cybermind-4b