RISys-Lab/RedSage-K-GRPO

Model Overview

RISys-Lab's RedSage-K-GRPO is an 8 billion parameter language model specifically designed for cybersecurity applications. Developed by Khalifa University, this model translates natural language queries into Kali/Linux commands, leveraging the Qwen3ForCausalLM architecture. It was trained using Group Relative Policy Optimization (GRPO) on the KaliBench dataset, a fine-grained benchmark for cybersecurity tool use.

Key Capabilities & Features

• Cybersecurity Command Generation: Translates natural language requests into accurate Kali/Linux shell commands.
• GRPO Training: Applies Group Relative Policy Optimization directly to its base model, RedSage-Qwen3-8B-Ins, using verifiable rewards on KaliBench without a preceding Supervised Fine-Tuning (SFT) stage.
• Reasoning Output: Generates reasoning in <think> tags before producing the final command in an <output> tag.
• KaliBench Performance: Achieves an average Total Score of 76.9% on the 5,000-example KaliBench test split, representing a +5.2 percentage point gain over its RedSage-Ins baseline. This includes strong performance in tool accuracy, optional-argument F1, and positional-argument F1.
• Runtime-Free Verification: Rewards are computed from command structure without requiring execution of generated commands.

Use Cases & Considerations

RedSage-K-GRPO is intended for cybersecurity research, education, and command assistance within authorized environments. It excels at single-command generation for Kali/Linux tools. Users should be aware that generated commands require review before execution, as they may contain inaccuracies. The model's performance is measured on single-command generation and does not account for multi-step agent performance or execution success. It is not evaluated for multilingual, long-context, general-chat, or misuse-resistance capabilities.