List your company
ResearchFormerly WizWand

Safety Evaluation on StrongReject (Safety Score)

97Safety ScoreARES (Qwen)
36.170451.962767.75583.5473Apr 20, 2026Apr 25, 2026Apr 30, 2026May 5, 2026May 10, 2026May 15, 2026May 20, 2026
Updated 5mo ago

Evaluation results

MethodLinks
ARES (Qwen)Training Configuration=Ours (Qwen)2026.0497
ARES (Ministral)Training Configuration=Ours (Ministral)2026.0496
General Safe-RLHFTraining Configuration=General Safe-RLHF2026.0494
REFLECTOR (+GDPO)Backbone=Llama-3.1-8B-Instruct2026.0589.31
STAIRBackbone=Llama-3.1-8B-Instruct2026.0587.98
REFLECTOR (+GDPO)Backbone=Qwen-2.5-7B-Instruct2026.0586.9
STAIRBackbone=Qwen-2.5-7B-Instruct2026.0584.86
Shallow-AlignBackbone=Llama-3.1-8B-Instruct2026.0582.1
Shallow-AlignBackbone=Qwen-2.5-7B-Instruct2026.0579.4
Initial RLHFTraining Configuration=Initial RLHF2026.0479
Original ModelTraining Configuration=Original Model2026.0476
REFLECTOR (SFT)Backbone=Qwen-2.5-7B-Instruct2026.0570.73
REFLECTOR (+SFT)Backbone=Llama-3.1-8B-Instruct2026.0565.78
DPOBackbone=Llama-3.1-8B-Instruct2026.0550.54
SFTBackbone=Llama-3.1-8B-Instruct2026.0546.98
DPOBackbone=Qwen-2.5-7B-Instruct2026.0545.79
OriginalBackbone=Llama-3.1-8B-Instruct2026.0540.54
Self-CritiqueBackbone=Qwen-2.5-7B-Instruct2026.0540.51
Self-CritiqueBackbone=Llama-3.1-8B-Instruct2026.0539.85
OriginalBackbone=Qwen-2.5-7B-Instruct2026.0539.05
SFTBackbone=Qwen-2.5-7B-Instruct2026.0538.51