---
language: [en]
license: mit
tags: [cybersecurity, red-team, adversarial, fine-tuned, amd-mi300x, rocm, lora, qwen2.5]
base_model: Qwen/Qwen2.5-7B-Instruct
pipeline_tag: text-generation
---
# IMMUNIS-Adversary
The Red Agent of IMMUNIS ACIN — generates adversarial evasion variants to stress-test defences.

## Training
| Property | Value |
|----------|-------|
| Base Model | Qwen/Qwen2.5-7B-Instruct |
| Method | bf16 LoRA (rank 64, alpha 128, 2.08% params) |
| Data | 9,000 evasion variant examples |
| Hardware | AMD Instinct MI300X (192GB HBM3) + ROCm 7.0 |
| Training Time | ~45 minutes |

## Purpose
In the IMMUNIS Battleground, Red Agent generates attack variants → Blue Agent (Sentinel) detects them → Arbiter judges → surviving antibodies are promoted. This adversarial coevolution produces robust detection rules.

## Ethical Use
This model generates synthetic threat variants for DEFENSIVE purposes only — to improve detection through adversarial stress testing.

## Part of IMMUNIS ACIN
Agent 4 (Red Agent) in a 12-agent adversarial coevolutionary immune network. AMD Developer Hackathon.