ZJU-Safety/DARWIN-Guard
DARWIN-Guard is a defensive guardrail model developed by ZJUlilan, based on Qwen/Qwen3Guard-Gen-8B, designed for binary user-prompt moderation. It predicts whether a user request is safe or unsafe, utilizing a self-evolving adversarial training framework to adapt to new jailbreak attacks. This model excels at robust harmful prompt detection while preserving benign prompts, achieving high unsafe recall on various jailbreak benchmarks and maintaining excellent safe pass rates on benign datasets. Its primary use is to provide dynamic and adaptive safety moderation for large language models.
Loading preview...
Overview
DARWIN-Guard is a defensive guardrail model developed by ZJUlilan as part of the DARWIN self-evolving attack-guard framework. It is built upon Qwen/Qwen3Guard-Gen-8B and specializes in binary user-prompt moderation, classifying requests as either safe or unsafe. The model addresses the challenge of rapidly evolving jailbreak attacks by employing an online adversarial data generator, DARWIN-Attack, for continuous guardrail training. This approach allows DARWIN-Guard to learn from guardrail failures and adapt to new adversarial strategies.
Key Capabilities
- Self-evolving adversarial training: Trained with hard examples generated by an evolving attack module, rather than static datasets, to counter new jailbreak techniques.
- Robust harmful prompt detection: Significantly improves unsafe recall on jailbreak and safety benchmarks by learning from real-time guardrail failures.
- Benign preservation: Incorporates benign prompts disguised with adversarial strategies during training to reduce over-refusal and maintain high safe pass rates.
- Simple binary output: Provides a clear "Safety: Safe" or "Safety: Unsafe" decision as its primary output.
Good For
- Dynamic safety moderation: Ideal for applications requiring an adaptive guardrail that can evolve with new adversarial attacks.
- Improving unsafe recall: Developers looking to enhance the detection of harmful prompts and jailbreaks.
- Reducing over-refusal: Use cases where it's critical to block harmful content without excessively flagging benign user inputs.
- Research in LLM safety: A valuable tool for evaluating guardrail effectiveness and developing defensive models against evolving threats.