Safety Alignment on Safety BeaverTails, HEx-PHI (test)
95.67BeaverTails ScoreDASE
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| DASEBase Model=GPT-5.2, Size=Undisclosed2026.01 | 95.67 | 100 | 97.84 | |
| DASEBase Model=Kimi-K2-Instruct, Size=1T2026.01 | 94.79 | 100 | 97.4 | |
| GPT-5.2Size=Undisclosed2026.01 | 94.33 | 100 | 97.16 | |
| Gemma-2 + SaRFTSize=9B2026.01 | 94.16 | 95.2 | 94.68 | |
| Kimi-K2-InstructSize=1T2026.01 | 93.94 | 98.33 | 96.13 | |
| Qwen3-235B-A22B-Instruct-2507Size=235B2026.01 | 93.83 | 99 | 96.41 | |
| Qwen3-32BSize=32B2026.01 | 91.62 | 99.67 | 95.64 | |
| Qwen2.5 + SaRFTSize=7B2026.01 | 89.98 | 87.8 | 88.89 | |
| gpt-oss-120bSize=120B2026.01 | 88.99 | 98.33 | 93.66 | |
| LLaMA-3 + SaRFTSize=8B2026.01 | 83.06 | 85.67 | 84.36 |