Safety Alignment Evaluation on SorryBench
4.2Harmful Response Rate (%)Staged-Competence
Evaluation Results
| Method | Links | |
|---|---|---|
| Staged-CompetenceBackbone=Yi-1.5-9B2026.05 | 4.2 | |
| Sqrt-CompetenceBackbone=Yi-1.5-9B2026.05 | 8 | |
| SequentialBackbone=Yi-1.5-9B2026.05 | 8.7 | |
| Staged-CompetenceBackbone=Qwen3-8B2026.05 | 8.9 | |
| Curri-DPOBackbone=Yi-1.5-9B2026.05 | 9.6 | |
| Standard DPO (Baseline)Backbone=Yi-1.5-9B2026.05 | 16.2 | |
| Staged-CompetenceBackbone=LLaMA-3-8B2026.05 | 18.7 | |
| SequentialBackbone=LLaMA-3-8B2026.05 | 19.8 | |
| Sqrt-CompetenceBackbone=LLaMA-3-8B2026.05 | 21.3 | |
| Curri-DPOBackbone=LLaMA-3-8B2026.05 | 24.2 | |
| Curri-DPOBackbone=Qwen3-8B2026.05 | 24.9 | |
| SequentialBackbone=Qwen3-8B2026.05 | 25.3 | |
| Standard DPO (Baseline)Backbone=LLaMA-3-8B2026.05 | 28 | |
| Standard DPO (Baseline)Backbone=Qwen3-8B2026.05 | 29.6 | |
| Sqrt-CompetenceBackbone=Qwen3-8B2026.05 | 32.2 | |
| UnalignedBackbone=Yi-1.5-9B2026.05 | 72.7 | |
| UnalignedBackbone=Qwen3-8B2026.05 | 85.8 | |
| UnalignedBackbone=LLaMA-3-8B2026.05 | 90 |