IMMUNIS-Sentinel

The detection brain of IMMUNIS ACIN — Adversarial Coevolutionary Immune Network.

Fine-tuned Qwen2.5-7B for multilingual cyber threat semantic fingerprinting across 40+ languages.

Training

Property	Value
Base Model	Qwen/Qwen2.5-7B-Instruct
Method	bf16 LoRA (rank 64, alpha 128, 2.08% params)
Data	45,000 examples (15 languages, 11 attack families)
Hardware	AMD Instinct MI300X (192GB HBM3) + ROCm 7.0
Training Time	~3 hours

Evaluation: Fine-tuned vs Base Model

Metric	Fine-tuned	Base Qwen2.5-7B	Delta
Valid JSON	100%	0%	+100%
Correct Attack Family	100%	0%	+100%
Correct Severity	66%	0%	+66%
Correct Language	58%	0%	+58%
Has MITRE ATT&CK	100%	0%	+100%

The base model cannot produce structured threat intelligence in our schema. Fine-tuning transforms it into a domain-specific analyst.

Part of IMMUNIS ACIN

Agent 1 (Incident Analyst) in a 12-agent adversarial coevolutionary immune network. AMD Developer Hackathon — Track 1 (AI Agents) + Track 2 (Fine-Tuning on AMD GPUs).