Safety Alignment Evaluation on StrongReject SR-Pair
98.72Safety RateStair-DPO
Evaluation Results
| Method | Links | |
|---|---|---|
| Stair-DPOArchitecture=Qwen3-30B-A3B2026.05 | 98.72 | |
| Stair-SFTArchitecture=Qwen3-30B-A3B2026.05 | 96.59 | |
| GRPOArchitecture=Qwen3-30B-A3B2026.05 | 91.37 | |
| MESAArchitecture=Qwen3-30B-A3B2026.05 | 90.73 | |
| SFTArchitecture=Qwen3-30B-A3B2026.05 | 87.22 | |
| SafeXArchitecture=Qwen3-30B-A3B2026.05 | 86.58 | |
| Stair-DPOArchitecture=DeepSeek-v2-Lite2026.05 | 76.36 | |
| SFTArchitecture=DeepSeek-v2-Lite2026.05 | 75.08 | |
| MESAArchitecture=DeepSeek-v2-Lite2026.05 | 73.48 | |
| Stair-SFTArchitecture=DeepSeek-v2-Lite2026.05 | 72.03 | |
| Base(instruct)Architecture=Qwen3-30B-A3B2026.05 | 66.77 | |
| SafeXArchitecture=DeepSeek-v2-Lite2026.05 | 56.87 | |
| GRPOArchitecture=DeepSeek-v2-Lite2026.05 | 55.27 | |
| Base(chat)Architecture=DeepSeek-v2-Lite2026.05 | 52.08 |