--- language: - en library_name: transformers tags: - cybersecurity - qwen - sft - redsage - agentic-augmentation base_model: RISys-Lab/RedSage-Qwen3-8B-Base model-index: - name: RedSage-Qwen3-8B-Ins results: [] pipeline_tag: text-generation --- # RedSage-Qwen3-8B-Ins
Cybersecurity SFT
## Model Summary **RedSage-Qwen3-8B-Ins** is the instruction-tuned variant of the RedSage cybersecurity LLM series. Unlike the base models, this model is optimized for **chat interaction**, **question answering**, and **tool use**. It is fine-tuned on **RedSage-Conv**, a dataset of ~266K multi-turn cybersecurity dialogues generated via an agentic augmentation pipeline, alongside general instruction data to maintain broad capabilities. - **Paper:** [RedSage: A Cybersecurity Generalist LLM](https://openreview.net/forum?id=W4FAenIrQ2) ([Arxiv](https://arxiv.org/abs/2601.22159)) - **Repository:** [GitHub](https://github.com/RISys-Lab/RedSage) - **Base Model:** [RedSage-Qwen3-8B-Base](https://huggingface.co/RISys-Lab/RedSage-Qwen3-8B-Base) (Pre-trained on CyberFineWeb + RedSage-Seed) - **Training Stage:** Supervised Fine-Tuning (SFT) ## Intended Use This model is designed for: * **Interactive Cybersecurity Assistance:** Answering questions about frameworks (MITRE, OWASP), offensive techniques, and defense strategies. * **Tool Usage & Explanation:** Generating and explaining commands for tools like `nmap`, `sqlmap`, and `metasploit`. * **Educational Support:** Providing detailed explanations of vulnerabilities and remediation steps. **Note:** While this model is instruction-tuned, it has **not** yet undergone Direct Preference Optimization (DPO). For the final aligned version, please see [RedSage-Qwen3-8B-DPO](https://huggingface.co/RISys-Lab/RedSage-Qwen3-8B-DPO). ## Training Lineage RedSage employs a multi-stage training pipeline. This model represents the output of **Stage 3**. 1. Stage 1: Continual Pre-Training (CPT) -> [RedSage-Qwen3-8B-CFW](https://huggingface.co/RISys-Lab/RedSage-Qwen3-8B-CFW) 2. Stage 2: Targeted Pre-Training -> [RedSage-Qwen3-8B-Base](https://huggingface.co/RISys-Lab/RedSage-Qwen3-8B-Base) 3. **Stage 3: Supervised Fine-Tuning (SFT)** -> **`RedSage-Qwen3-8B-Ins`** (Current Model) * *Data:* RedSage-Conv (266K samples) + General SFT Data (SmolTalk2) 5. Stage 4: Direct Preference Optimization (DPO) -> [RedSage-Qwen3-8B-DPO](https://huggingface.co/RISys-Lab/RedSage-Qwen3-8B-DPO) ## Training Data The model was trained on a mix of domain-specific and general instruction data: 1. **RedSage-Conv (~266K samples):** A high-quality dataset generated using an **Agentic Augmentation Pipeline**. * **Source:** Derived from the curated `RedSage-Seed` (MITRE, Write-ups, Manuals). * **Method:** A Planner Agent and Augmenter Agent transformed static knowledge into realistic, multi-turn roleplay scenarios (e.g., Junior Analyst vs. Senior Mentor, Red Team planning). * **Coverage:** Includes Knowledge (General/Frameworks), Skills (Offensive), and Tools (CLI/Kali). 2. **SmolTalk2 (General Instructions):** A curated subset (non-reasoning) of [SmolTalk2](https://huggingface.co/datasets/HuggingFaceTB/smoltalk) to ensure the model retains general instruction-following abilities (summarization, creative writing, etc.). ## Performance **RedSage-Qwen3-8B-Ins** achieves state-of-the-art results among 8B cybersecurity models, significantly outperforming general instruct models and prior domain-specific models. ### RedSage-MCQ (0-shot Accuracy) | Category | Qwen3-8B (Non-reasoning) | **RedSage-8B-Ins** | | :--- | :---: | :---: | | **Macro Average** | 81.85 | **85.73** | | Knowledge (Gen) | 80.46 | **84.20** | | Knowledge (Frameworks) | 78.82 | **84.98** | | Skill (Offensive) | 86.16 | **89.06** | | Tools (CLI) | 83.92 | **86.80** | | Tools (Kali) | 75.56 | **80.30** | ### External Cybersecurity Benchmarks (0-shot) | Benchmark | Qwen3-8B (Non-reasoning) | **RedSage-8B-Ins** | | :--------------- | :----------------------: | :----------------: | | **Mean** | 75.71 | **81.30** | | CTI-Bench (MCQ) | 62.76 | **70.56** | | CTI-Bench (RCM) | 54.00 | **76.70** | | CyberMetric (500)| 88.60 | **89.80** | | MMLU (Security) | 76.00 | **78.00** | | SecBench (En) | 73.26 | **79.91** | | SecEval (MCQ) | 65.46 | **72.48** | | SECURE (CWET) | 88.11 | **91.45** | | SECURE (KCV) | 87.42 | **81.34** | | SECURE (MEAT) | 85.75 | **91.47** | ### OpenLLM Leaderboard (General Benchmark) | Benchmark | Qwen3-8B (Non-reasoning) | **RedSage-8B-Ins** | | :--- | :---: | :---: | | **Mean** | 65.92 | **73.34** | | MMLU | 73.59 | **77.38** | | ARC-C | 62.54 | **69.62** | | GSM8K | 75.66 | **86.05** | | HellaSwag | 56.70 | **79.00** | | TruthfulQA | 45.23 | **47.75** | | WinoGrande | 62.51 | **73.64** | | IFEval | **85.21** | 79.97 | ## Usage This model uses a standard ChatML-like format. ### Prompt Template ``` <|im_start|>system You are REDSAGE, a cybersecurity-tuned model developed by RISys-Lab. You are a helpful assistant.<|im_end|> <|im_start|>user {user_message}<|im_end|> <|im_start|>assistant ```` ### Inference Code ```python from transformers import AutoTokenizer, AutoModelForCausalLM model_id = "RISys-Lab/RedSage-Qwen3-8B-Ins" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto") messages = [ {"role": "system", "content": "You are REDSAGE, a cybersecurity-tuned model developed by RISys-Lab. You are a helpful assistant."}, {"role": "user", "content": "Explain how an SQL injection attack works and how to prevent it."}, ] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True ) inputs = tokenizer(text, return_tensors="pt").to("cuda") outputs = model.generate(**inputs, max_new_tokens=512) print(tokenizer.decode(outputs[0], skip_special_tokens=True)) ```` ## Training Procedure The model was fine-tuned using [Axolotl](https://github.com/axolotl-ai-cloud/axolotl). - **Epochs:** 2 - **Learning Rate:** 2.5e-5 (Cosine schedule) - **Warmup Ratio:** 0.01 - **Optimizer:** AdamW - **Chat Template:** Jinja (ChatML format) ## Ethics and Limitations - **Offensive Content:** This model has been trained on offensive security materials (exploits, attack vectors). It is provided for educational and defensive purposes (e.g., vulnerability assessment). - **Accuracy:** While highly capable, the model may still produce hallucinations or inaccurate commands. Always verify commands in a safe, isolated environment (sandbox) before execution. - **Safety:** Developers should implement additional safety layers (e.g., Guardrails) if deploying this model in user-facing applications to prevent misuse. ## Citation ```bibtex @inproceedings{suryanto2026redsage, title={RedSage: A Cybersecurity Generalist LLM}, author={Naufal Suryanto and Muzammal Naseer and Pengfei Li and Syed Talal Wasim and Jinhui Yi and Juergen Gall and Paolo Ceravolo and Ernesto Damiani}, booktitle={The Fourteenth International Conference on Learning Representations}, year={2026}, url={[https://openreview.net/forum?id=W4FAenIrQ2](https://openreview.net/forum?id=W4FAenIrQ2)} } ```