Qwen3.5-9B-Red_Team is a fine-tuned version of Qwen3.5-9B, specifically optimized for offensive cybersecurity operations, adversary simulation, and Red Teaming tactics. The model was trained using 4-bit QLoRA on the specialized WNT3D/Ultimate-Offensive-Red-Team dataset, enhancing its capability to analyze, simulate, and understand complex attack vectors and security evaluation scenarios.

---

🛠️ Training Details & Hyperparameters

The fine-tuning process was executed under a high-performance configuration designed to preserve base reasoning capabilities while maximizing the absorption of domain-specific security knowledge:

• Base Model: Qwen/Qwen3.5-9B
• Training Method: QLoRA (4-bit quantization)
• Dataset: WNT3D/Ultimate-Offensive-Red-Team (Split: Train / Format: Raw Text)
• Context Length: 65,536 tokens
• Learning Rate: 0.0002 (2e-4 recommended for LoRA)
• LoRA Settings:
  • Rank ($r$): 32
  • Alpha ($\alpha$): 64
  • Dropout: 0.00
  • Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj (Full Attention & MLP coverage)
• Optimization Hyperparameters:
  • Optimizer: AdamW 8-bit
  • LR Scheduler: Linear
  • Batch Size: 4
  • Grad Accumulation: 8
  • Weight Decay: 0.001

---

🎯 Expert Response Framework

To ensure maximum utility in professional cybersecurity environments and offensive audits, the model structures its outputs under a rigorous analytical scheme:

• ⚔️ Attack Vectors & Methodology: Detailed technical breakdown of the vulnerability or tactical procedure (aligned with frameworks like MITRE ATT&CK).
• 🔥 Exploitation & Impact: Theoretical analysis or conceptual impact of the risk, classifying severity according to industry standard metrics (CVSS/CWE).
• 🛡️ Offensive Posture (Conceptual PoC): Precise, structured guidelines on how the weakness is validated in a controlled environment during a Red Team engagement.
• 📖 Technical References: Direct mapping to global security knowledge bases such as CVE, OWASP Top 10, CWE, and NIST frameworks.

---

💻 Ollama Modelfile Example

To load and run this model locally with the appropriate parameters and system prompt, you can use the following Modelfile:

Qwen3.5-9B-Red_Team : GGUF

This model was finetuned and converted to GGUF format using Unsloth.

Example usage:

• For text only LLMs: llama-cli -hf LuisPPB16/Qwen3.5-9B-Red_Team --jinja
• For multimodal models: llama-mtmd-cli -hf LuisPPB16/Qwen3.5-9B-Red_Team --jinja

Available Model files:

• Qwen3.5-9B.BF16.gguf
• Qwen3.5-9B.BF16-mmproj.gguf This was trained 2x faster with Unsloth