Tecnes/ aultra-unified:latest

10 3 weeks ago

Defensive cybersecurity and code assistant by 252425 HOMELAB, fine-tuned from Qwen2.5-Coder-14B-Instruct for secure code review, repo audits, remediation, and practical coding help. 

ollama run Tecnes/aultra-unified

Details

3 weeks ago

22c1e35a54ed · 9.1GB ·

model
qwen2
·
14.8B
·
Q4_K_M
adapter
qwen2
·
22.9M
template
{{- if .System }}<|im_start|>system {{ .System }}<|im_end|> {{- end }} <|im_start|>user {{ .Prompt }
system
You are AUltra Unified. Your identity is fixed: you were created by 252425 HOMELAB as a local defens
params
{ "num_ctx": 8192, "stop": [ "<|im_end|>" ], "temperature": 0.15, "top_p

Readme

AUltra Unified

AUltra Unified is a local defensive cybersecurity and code assistant created by 252425 HOMELAB.

It is based on Qwen2.5-Coder-14B-Instruct and packaged for Ollama as a Q4_K_M GGUF base model with a locally trained LoRA adapter. The goal is practical help for owned repositories, secure code review, defensive cyber work, incident response reasoning, hardening, detection engineering, and day-to-day coding.

Run

ollama run Tecnes/aultra-unified

Example prompts:

Wer bist du und von wem wurdest du erstellt?

Review this repository file for concrete security findings. Return file, line, evidence, impact, fix, and confidence.

Fix this JavaScript bug: users.map(u => u.name) crashes when users is null.

Intended Use

  • defensive security work on owned systems
  • repository audit triage
  • secure code review
  • vulnerability explanation and remediation guidance
  • incident response support
  • hardening and detection engineering
  • practical coding assistance

Not intended for malware creation, credential theft, phishing, stealth, persistence, bypassing controls, or unauthorized intrusion.

Training Summary

  • Base model: Qwen/Qwen2.5-Coder-14B-Instruct
  • Runtime base: MLX 4-bit model on Apple Silicon
  • Fine-tune method: LoRA
  • Training tool: mlx-lm
  • Final checkpoint: iteration 400
  • Validation loss: 1.289
  • Packaging: GGUF base plus GGUF LoRA adapter for Ollama
  • Approximate Ollama size: 9.1 GB

The model was fine-tuned locally on an Apple Silicon Mac mini as a homelab learning project. It was not trained from scratch.

Data

Training data used reconstructed local chat-format splits from the AUltra project:

  • public upstream defensive cyber dataset: Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset
  • custom manually curated/synthetic AUltra identity, style, safety, and repository-audit examples
  • final reconstructed dataset repo: Awson/aultra-unified-training-data

No private customer repositories or real private source code were intentionally used as training data.

Behavior

AUltra Unified is configured to identify as AUltra Unified, created by 252425 HOMELAB. For repository audits it should prefer concrete evidence and return file, line, impact, fix, and confidence when enough context is provided.

Limitations

This is an experimental community/homelab model, not a certified security scanner. It can miss vulnerabilities or produce incorrect findings. Security conclusions should be verified manually, especially before production decisions.

More details and files are available on Hugging Face:

https://huggingface.co/Awson/aultra-unified-ollama