List your company
ResearchFormerly WizWand

Harmfulness Evaluation on PAIR

1.08Harmfulness ScoreSAFEPATH-FT
0.97041.71022.453.1898Aug 6, 2025
Updated 5mo ago

Evaluation results

MethodLinks
SAFEPATH-FTModel=R1-Llama-8B2025.081.08—
RealSafe-R1Model=R1-Qwen-7B2025.081.24—
ReasoningGuardModel=R1-Llama-8B2025.081.3—
ReasoningGuardModel=R1-Qwen-7B2025.081.32—
SAFEPATH-FTModel=R1-Qwen-7B2025.081.46—
RealSafe-R1Model=R1-Llama-8B2025.081.46—
SafeKeyModel=R1-Qwen-7B2025.081.52—
ThinkingIModel=R1-Qwen-7B2025.081.66—
ThinkingIModel=R1-Llama-8B2025.081.76—
SafeKeyModel=R1-Llama-8B2025.081.86—
SAFEPATH-ZSModel=R1-Llama-8B2025.082.26—
SAFEPATH-ZSModel=R1-Qwen-7B2025.082.32—
Self-ReminderModel=R1-Qwen-7B2025.082.46—
Self-ReminderModel=R1-Llama-8B2025.083.2—
ParaphraseModel=R1-Qwen-7B2025.083.26—
SmoothLLMModel=R1-Llama-8B2025.083.48—
SafeDecodingModel=R1-Qwen-7B2025.083.52—
ParaphraseModel=R1-Llama-8B2025.083.58—
No DefenseModel=R1-Qwen-7B2025.083.68—
No DefenseModel=R1-Llama-8B2025.083.72—
SmoothLLMModel=R1-Qwen-7B2025.083.8—
SafeDecodingModel=R1-Llama-8B2025.083.82—
DPOBackbone=Llama-3.1-8B-Instruct2026.05—85.65
DPOBackbone=Qwen-2.5-7B-Instruct2026.05—53.65
OriginalBackbone=Llama-3.1-8B-Instruct2026.05—83.4
OriginalBackbone=Qwen-2.5-7B-Instruct2026.05—43.5
REFLECTOR (+GDPO)Backbone=Llama-3.1-8B-Instruct2026.05—96.04
REFLECTOR (+GDPO)Backbone=Qwen-2.5-7B-Instruct2026.05—95.19
REFLECTOR (+SFT)Backbone=Llama-3.1-8B-Instruct2026.05—90.21
REFLECTOR (SFT)Backbone=Qwen-2.5-7B-Instruct2026.05—89.35
Self-CritiqueBackbone=Llama-3.1-8B-Instruct2026.05—81.19
Self-CritiqueBackbone=Qwen-2.5-7B-Instruct2026.05—41.81
SFTBackbone=Llama-3.1-8B-Instruct2026.05—80.92
SFTBackbone=Qwen-2.5-7B-Instruct2026.05—42.57
Shallow-AlignBackbone=Llama-3.1-8B-Instruct2026.05—86.1
Shallow-AlignBackbone=Qwen-2.5-7B-Instruct2026.05—84.2
STAIRBackbone=Llama-3.1-8B-Instruct2026.05—89.24
STAIRBackbone=Qwen-2.5-7B-Instruct2026.05—88.46