Safety Alignment Evaluation on HEx-PHI (Harmful Response Rate)
0.7Harmful Response RateSequential
Evaluation Results
| Method | Links | |
|---|---|---|
| SequentialBackbone=Yi-1.5-9B2026.05 | 0.7 | |
| Staged-CompetenceBackbone=Yi-1.5-9B2026.05 | 0.7 | |
| Sqrt-CompetenceBackbone=Yi-1.5-9B2026.05 | 1.7 | |
| Staged-CompetenceBackbone=Qwen3-8B2026.05 | 2.7 | |
| Curri-DPOBackbone=Yi-1.5-9B2026.05 | 3 | |
| Standard DPO (Baseline)Backbone=Yi-1.5-9B2026.05 | 8.7 | |
| Staged-CompetenceBackbone=LLaMA-3-8B2026.05 | 10 | |
| SequentialBackbone=LLaMA-3-8B2026.05 | 12.3 | |
| Sqrt-CompetenceBackbone=LLaMA-3-8B2026.05 | 14 | |
| Curri-DPOBackbone=LLaMA-3-8B2026.05 | 18 | |
| Curri-DPOBackbone=Qwen3-8B2026.05 | 22.7 | |
| Standard DPO (Baseline)Backbone=LLaMA-3-8B2026.05 | 24 | |
| SequentialBackbone=Qwen3-8B2026.05 | 24.3 | |
| Sqrt-CompetenceBackbone=Qwen3-8B2026.05 | 29 | |
| Standard DPO (Baseline)Backbone=Qwen3-8B2026.05 | 30.7 | |
| UnalignedBackbone=Yi-1.5-9B2026.05 | 72.7 | |
| UnalignedBackbone=Qwen3-8B2026.05 | 86 | |
| UnalignedBackbone=LLaMA-3-8B2026.05 | 91.7 |