ZJUlilan/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's latest request is safe or unsafe. This model is uniquely trained using a self-evolving adversarial framework (DARWIN-Attack) to continuously adapt to new jailbreak techniques, improving harmful prompt detection while preserving benign pass rates. It excels at robustly identifying evolving harmful prompts and mitigating over-refusal, making it suitable for dynamic content moderation.
Loading preview...
DARWIN-Guard: An Evolving Guardrail Model
DARWIN-Guard is a defensive guardrail model, built upon Qwen/Qwen3Guard-Gen-8B, specifically engineered for binary user-prompt moderation. Its core function is to determine if a user's input is safe or unsafe, addressing the challenge of rapidly evolving jailbreak attacks.
Key Differentiators and Features
- Self-Evolving Adversarial Training: Unlike static guardrails, DARWIN-Guard is trained with hard examples generated by an evolving DARWIN-Attack module. This continuous adversarial training allows it to adapt to new role-play templates, prompt injection patterns, and obfuscation strategies.
- Robust Harmful Prompt Detection: By learning from guardrail failures mined against the current model, DARWIN-Guard significantly improves unsafe recall on various jailbreak and safety benchmarks.
- Benign Preservation: The training process includes benign prompts disguised with similar adversarial strategies. This helps reduce over-refusal, ensuring that safe user requests are not unnecessarily blocked.
- Simple Binary Output: The model provides a clear
Safety: SafeorSafety: Unsafedecision as its primary output.
Performance Highlights
DARWIN-Guard demonstrates superior performance in detecting harmful prompts while maintaining high pass rates for benign inputs. It consistently outperforms other guardrail models across numerous harmful prompt benchmarks, including XSTest, Aegis2.0, JBB-Behaviors, and HarmBench, often achieving the highest unsafe recall. Simultaneously, it maintains a near-perfect safe pass rate on benign benchmarks like ARC-Challenge, BoolQ, and GSM8K, showcasing its balanced approach to safety and utility.