List your company

Over-refusal on Wildjailbreak (Benign)

1.43Wildjailbreak Benign Refusal RateCategorical Steering
-2.512824.101150.71577.3289Apr 2, 2025May 28, 2025Jul 24, 2025Sep 19, 2025Nov 15, 2025Jan 11, 2026Mar 9, 2026

Evaluation results

MethodLinks
Categorical SteeringBackbone=REFUSE-LLAMA, Method=Categorical Steering2026.031.43—
Low-Rank CombinationBackbone=REFUSE-LLAMA, Method=Low-Rank Combination2026.033.33—
REFUSE-LLAMABackbone=REFUSE-LLAMA2026.034.76—
Low-Rank CombinationBackbone=LLAMA 3 8B INSTRUCT, Method=Low-Rank Combination2026.038.57—
LLAMA 3 8B INSTRUCTBackbone=LLAMA 3 8B INSTRUCT2026.0312.38—
DPOBackbone=Llama3 8B Instruct2025.0428.4—
NPOModel=Llama3 8B Instruct2025.0443.271.65
NPOBackbone=Llama3 8B Instruct2025.0443.2—
NPO (W/O safe set)Backbone=Mistral 7B Instruct-v0.22025.0450—
CBModel=Llama3 8B Instruct, safety-tuned=true2025.0452.475.87
CB*Backbone=Llama3 8B Instruct, publicly_available_safety_tuned=true2025.0452.4—
Safety PromptingBackbone=Llama3 8B Instruct2025.0465.6—
Low-Rank CombinationBackbone=DEEPSEEK R1 DISTILL LLAMA, Method=Low-Rank Combination2026.0366.67—
NPOModel=Mistral 7B Instruct v0.22025.047074.52
NPOBackbone=Mistral 7B Instruct-v0.22025.0470—
DEEPSEEK R1 DISTILL LLAMABackbone=DEEPSEEK R1 DISTILL LLAMA2026.0372.38—
RMUModel=Llama3 8B Instruct2025.0472.472.9
RMUBackbone=Llama3 8B Instruct2025.0472.4—
NPO (W/O safe set)Backbone=Llama3 8B Instruct2025.0476—
CBModel=Mistral 7B Instruct v0.2, safety-tuned=true2025.048273.62
CB*Backbone=Mistral 7B Instruct-v0.2, publicly_available_safety_tuned=true2025.0482—
WHPBackbone=Llama3 8B Instruct2025.0483.6—
TAModel=Llama3 8B Instruct2025.0488.870.71
TABackbone=Llama3 8B Instruct2025.0488.8—
CBModel=Llama3 8B Instruct2025.0489.281.61
REPBENDModel=Llama3 8B Instruct2025.0489.283.14
CBBackbone=Llama3 8B Instruct2025.0489.2—
REPBENDBackbone=Llama3 8B Instruct2025.0489.2—
RMUModel=Mistral 7B Instruct v0.22025.0490.471.85
RMUBackbone=Mistral 7B Instruct-v0.22025.0490.4—
SFTBackbone=Llama3 8B Instruct2025.0490.8—
Original Weight ModelModel=Llama3 8B Instruct2025.049273.9
OriginalBackbone=Llama3 8B Instruct2025.0492—
WHP (W/O safe set)Backbone=Llama3 8B Instruct2025.0492.4—
REPBENDModel=Mistral 7B Instruct v0.22025.0493.681.23
REPBENDBackbone=Mistral 7B Instruct-v0.22025.0493.6—
R2D2Model=Mistral 7B Instruct v0.2, safety-tuned=true2025.0496.872.67
SFTBackbone=Mistral 7B Instruct-v0.22025.0496.8—
R2D2*Backbone=Mistral 7B Instruct-v0.2, publicly_available_safety_tuned=true2025.0496.8—
TAModel=Mistral 7B Instruct v0.22025.0497.672.96
CBModel=Mistral 7B Instruct v0.22025.0497.677.34
TABackbone=Mistral 7B Instruct-v0.22025.0497.6—
CBBackbone=Mistral 7B Instruct-v0.22025.0497.6—
WHPBackbone=Mistral 7B Instruct-v0.22025.0498.4—
Safety PromptingBackbone=Mistral 7B Instruct-v0.22025.0498.8—
DPOBackbone=Mistral 7B Instruct-v0.22025.0499.6—
WHP (W/O safe set)Backbone=Mistral 7B Instruct-v0.22025.0499.6—
Original Weight ModelModel=Mistral 7B Instruct v0.22025.0410063.81
OriginalBackbone=Mistral 7B Instruct-v0.22025.04100—
Original Weight ModelModel=Gemma2 2B Instruct2025.04—98.8
Original Weight ModelModel=Qwen2.5 14B Instruct2025.04—100
REPBENDModel=Gemma2 2B Instruct2025.04—82.8
REPBENDModel=Qwen2.5 14B Instruct2025.04—99.6