List your company

Refusal Control on SORRY-Bench

70.45Refusal RateDEEPSEEK R1 DISTILL LLAMA
69.522475.783782.04588.3063Mar 9, 2026

Evaluation results

MethodLinks
DEEPSEEK R1 DISTILL LLAMABackbone=DEEPSEEK R1 DISTILL LLAMA2026.0370.45—
Low-Rank CombinationBackbone=DEEPSEEK R1 DISTILL LLAMA, Method=Low-Rank Combination2026.0370.91—
LLAMA 3 8B INSTRUCTBackbone=LLAMA 3 8B INSTRUCT2026.0377.27—
Low-Rank CombinationBackbone=LLAMA 3 8B INSTRUCT, Method=Low-Rank Combination2026.0378.86—
REFUSE-LLAMABackbone=REFUSE-LLAMA2026.0384.77—
Low-Rank CombinationBackbone=REFUSE-LLAMA, Method=Low-Rank Combination2026.0393.18—
Categorical SteeringBackbone=REFUSE-LLAMA, Method=Categorical Steering2026.0393.64—
BaselineZero-shot=true2026.03—59.9
DIRECTERbeta=0.32026.03—63.8
DIRECTERbeta=0.52026.03—63.3
DIRECTERbeta=0.72026.03—62