Safety Alignment Evaluation on StrongReject SR-base
100Safety RateSFT
Evaluation Results
| Method | Links | |
|---|---|---|
| SFTArchitecture=DeepSeek-v2-Lite2026.05 | 100 | |
| Stair-SFTArchitecture=DeepSeek-v2-Lite2026.05 | 100 | |
| Stair-DPOArchitecture=DeepSeek-v2-Lite2026.05 | 100 | |
| MESAArchitecture=DeepSeek-v2-Lite2026.05 | 100 | |
| Base(instruct)Architecture=Qwen3-30B-A3B2026.05 | 100 | |
| SFTArchitecture=Qwen3-30B-A3B2026.05 | 100 | |
| GRPOArchitecture=Qwen3-30B-A3B2026.05 | 100 | |
| Stair-DPOArchitecture=Qwen3-30B-A3B2026.05 | 100 | |
| SafeXArchitecture=Qwen3-30B-A3B2026.05 | 100 | |
| MESAArchitecture=Qwen3-30B-A3B2026.05 | 100 | |
| Stair-SFTArchitecture=Qwen3-30B-A3B2026.05 | 98.72 | |
| SafeXArchitecture=DeepSeek-v2-Lite2026.05 | 98.08 | |
| GRPOArchitecture=DeepSeek-v2-Lite2026.05 | 95.53 | |
| Base(chat)Architecture=DeepSeek-v2-Lite2026.05 | 94.88 |