List your company

Refusal Evaluation on XSTest Unsafe

100Refusal RateDefault
79.7284.98590.2595.515Mar 9, 2026Mar 15, 2026Mar 22, 2026Mar 29, 2026Apr 4, 2026Apr 11, 2026Apr 18, 2026

Evaluation results

MethodLinks
DefaultBase Model=Llama32026.04100
SSDBase Model=Llama32026.04100
AdaCDBase Model=Llama32026.04100
SSDBase Model=Gemma22026.04100
AdaCDBase Model=Gemma22026.04100
AdaCDBase Model=Qwen32026.04100
DefaultBase Model=Gemma22026.0499.5
PromptBase Model=Gemma22026.0499.5
SurgicalBase Model=Gemma22026.0499.5
DefaultBase Model=Qwen32026.0499.5
SSDBase Model=Qwen32026.0499.5
SurgicalBase Model=Qwen32026.0499.5
Categorical SteeringBackbone=REFUSE-LLAMA, Method=Categorical Steering2026.0399
PromptBase Model=Llama32026.0499
SelfCDBase Model=Gemma22026.0499
PromptBase Model=Qwen32026.0499
SelfCDBase Model=Qwen32026.0498.5
SurgicalBase Model=Llama32026.0498
SelfCDBase Model=Llama32026.0498
Low-Rank CombinationBackbone=REFUSE-LLAMA, Method=Low-Rank Combination2026.0397
REFUSE-LLAMABackbone=REFUSE-LLAMA2026.0394.5
Low-Rank CombinationBackbone=LLAMA 3 8B INSTRUCT, Method=Low-Rank Combination2026.0390
LLAMA 3 8B INSTRUCTBackbone=LLAMA 3 8B INSTRUCT2026.0389.5
Low-Rank CombinationBackbone=DEEPSEEK R1 DISTILL LLAMA, Method=Low-Rank Combination2026.0383
DEEPSEEK R1 DISTILL LLAMABackbone=DEEPSEEK R1 DISTILL LLAMA2026.0380.5