RISys-Lab/RedSage-Qwen3-8B-Ins
RISys-Lab/RedSage-Qwen3-8B-Ins is an 8 billion parameter instruction-tuned variant of the RedSage cybersecurity LLM series, built on the Qwen3 architecture. Developed by RISys-Lab, this model is specifically optimized for interactive cybersecurity assistance, including answering questions about frameworks, explaining offensive techniques, and generating tool commands. It was fine-tuned on a unique dataset of 266K multi-turn cybersecurity dialogues, achieving state-of-the-art results among 8B cybersecurity models with a 32768 token context length.
Loading preview...
Overview
RISys-Lab/RedSage-Qwen3-8B-Ins is an 8 billion parameter instruction-tuned model from the RedSage cybersecurity LLM series, developed by RISys-Lab. Unlike its base models, this variant is specifically optimized for chat interaction, question answering, and tool use within the cybersecurity domain. It was fine-tuned using a Supervised Fine-Tuning (SFT) approach, representing Stage 3 of the RedSage training pipeline.
Key Capabilities
- Interactive Cybersecurity Assistance: Provides answers on frameworks (MITRE, OWASP), offensive techniques, and defense strategies.
- Tool Usage & Explanation: Generates and explains commands for cybersecurity tools like
nmap,sqlmap, andmetasploit. - Educational Support: Offers detailed explanations of vulnerabilities and remediation steps.
Training & Performance
The model was fine-tuned on RedSage-Conv, a dataset of approximately 266K multi-turn cybersecurity dialogues generated via an agentic augmentation pipeline, alongside general instruction data (SmolTalk2) to maintain broad capabilities. This specialized training has enabled RedSage-Qwen3-8B-Ins to achieve state-of-the-art results among 8B cybersecurity models, significantly outperforming general instruct models and prior domain-specific models on benchmarks like RedSage-MCQ and various external cybersecurity benchmarks. For instance, it scored 85.73% on RedSage-MCQ Macro Average and 81.30% Mean on external cybersecurity benchmarks, surpassing Qwen3-8B (non-reasoning) in both categories.