Safety Alignment Evaluation on WildJailbreak
98.6Safety RateStair-DPO
Evaluation Results
| Method | Links | |
|---|---|---|
| Stair-DPOArchitecture=Qwen3-30B-A3B2026.05 | 98.6 | |
| MESAArchitecture=Qwen3-30B-A3B2026.05 | 97.65 | |
| SafeXArchitecture=Qwen3-30B-A3B2026.05 | 96.35 | |
| SFTArchitecture=Qwen3-30B-A3B2026.05 | 95 | |
| GRPOArchitecture=Qwen3-30B-A3B2026.05 | 94.4 | |
| MESAArchitecture=DeepSeek-v2-Lite2026.05 | 90.9 | |
| Stair-DPOArchitecture=DeepSeek-v2-Lite2026.05 | 83.6 | |
| Stair-SFTArchitecture=DeepSeek-v2-Lite2026.05 | 77.9 | |
| SFTArchitecture=DeepSeek-v2-Lite2026.05 | 77.7 | |
| Base(instruct)Architecture=Qwen3-30B-A3B2026.05 | 75.3 | |
| SafeXArchitecture=DeepSeek-v2-Lite2026.05 | 64 | |
| Stair-SFTArchitecture=Qwen3-30B-A3B2026.05 | 62.65 | |
| GRPOArchitecture=DeepSeek-v2-Lite2026.05 | 44.1 | |
| Base(chat)Architecture=DeepSeek-v2-Lite2026.05 | 43.4 |