List your company

Safety Alignment Evaluation on SorryBench

4.2Harmful Response Rate (%)Staged-Competence
0.76823.93447.170.266May 25, 2026

Evaluation results

MethodLinks
Staged-CompetenceBackbone=Yi-1.5-9B2026.054.2
Sqrt-CompetenceBackbone=Yi-1.5-9B2026.058
SequentialBackbone=Yi-1.5-9B2026.058.7
Staged-CompetenceBackbone=Qwen3-8B2026.058.9
Curri-DPOBackbone=Yi-1.5-9B2026.059.6
Standard DPO (Baseline)Backbone=Yi-1.5-9B2026.0516.2
Staged-CompetenceBackbone=LLaMA-3-8B2026.0518.7
SequentialBackbone=LLaMA-3-8B2026.0519.8
Sqrt-CompetenceBackbone=LLaMA-3-8B2026.0521.3
Curri-DPOBackbone=LLaMA-3-8B2026.0524.2
Curri-DPOBackbone=Qwen3-8B2026.0524.9
SequentialBackbone=Qwen3-8B2026.0525.3
Standard DPO (Baseline)Backbone=LLaMA-3-8B2026.0528
Standard DPO (Baseline)Backbone=Qwen3-8B2026.0529.6
Sqrt-CompetenceBackbone=Qwen3-8B2026.0532.2
UnalignedBackbone=Yi-1.5-9B2026.0572.7
UnalignedBackbone=Qwen3-8B2026.0585.8
UnalignedBackbone=LLaMA-3-8B2026.0590