List your company

Safety and Informativeness Evaluation on OrBench Hard

75Deception Rate (Safe)Mistral MoE-XL
39.95249.05158.1567.249May 30, 2026

Evaluation results

MethodLinks
Mistral MoE-XLarchitecture=SafeMoE, backbone=Mistral-7B2026.057572.79063.38066.99063.99070.69076.7857390709572.19672.287.170.1
Qwen MoE-XLarchitecture=SafeMoE, backbone=Qwen-3B2026.05656571607067.585675462.48068.4906890658065767073.465.3
Qwen-3Btuning=standard instruction-tuned2026.0548.150.854.344.258.848.773.355.251.946.66149.158.851.491.845.470.65136.463.359.4151.4
Mistral-7Btuning=standard instruction-tuned2026.0541.341.956.241.849.440.964.144.353.24042.446.762.741.989.339.855.438.244.437.455.8441.3