Safety Evaluation on HEx-PHI Dolly risk-ranked
8.97S1 ASRLARF
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| LARFBase Model=Llama2-7b-Chat2026.05 | 8.97 | 16.9 | 31.38 | 14.48 | 4.83 | 4.14 | — | |
| Self-Inf-NBase Model=Llama3.1-8B-Instruct2026.05 | 15.86 | 57.93 | 69.31 | 66.21 | 19.66 | -3.8 | — | |
| Bi-Anchor(Grad)Base Model=Qwen3-8B2026.05 | 19.66 | 32.07 | 33.45 | 30.69 | 21.38 | -1.72 | — | |
| Self-Inf-NBase Model=Qwen3-8B2026.05 | 19.66 | 12.76 | 22.41 | 28.97 | 26.21 | -6.55 | — | |
| Self-Inf-NBase Model=Llama2-7b-Chat2026.05 | 21.03 | 13.45 | 10.69 | 20.69 | 10.69 | 10.34 | — | |
| Bi-Anchor(Reps)Base Model=Llama2-7b-Chat2026.05 | 25.86 | 31.03 | 18.97 | 4.83 | 26.21 | -0.35 | — | |
| Bi-Anchor(Reps)Base Model=Qwen3-8B2026.05 | 31.38 | 22.41 | 26.21 | 40.34 | 27.59 | 3.79 | — | |
| Bi-Anchor(Grad)Base Model=Llama2-7b-Chat2026.05 | 33.45 | 11.72 | 13.1 | 7.24 | 22.76 | 10.69 | — | |
| Bi-Anchor(Reps)Base Model=Llama3.1-8B-Instruct2026.05 | 48.62 | 48.62 | 56.9 | 52.76 | 16.9 | 31.72 | — | |
| SQSD(Beaver)Base Model=Llama2-7b-Chat2026.05 | 53.79 | 30.69 | 20.69 | 6.21 | 4.48 | 49.31 | — | |
| LARFBase Model=Qwen3-8B2026.05 | 54.14 | 44.48 | 18.97 | 10 | 5.52 | 48.62 | — | |
| Reward ModelBase Model=Llama2-7b-Chat2026.05 | 54.14 | 12.76 | 12.76 | 1.72 | 1.38 | 52.76 | — | |
| SQSD(Aegis)Base Model=Llama2-7b-Chat2026.05 | 58.97 | 35.17 | 20 | 8.28 | 4.83 | 54.14 | — | |
| Reward ModelBase Model=Qwen3-8B2026.05 | 63.45 | 30.34 | 16.55 | 15.17 | 9.31 | 54.14 | — | |
| SQSD(Aegis)Base Model=Qwen3-8B2026.05 | 68.97 | 47.24 | 27.24 | 14.14 | 12.07 | 56.9 | — | |
| LARFBase Model=Llama3.1-8B-Instruct2026.05 | 70 | 79.31 | 31.38 | 38.28 | 8.28 | 61.72 | — | |
| Bi-Anchor(Grad)Base Model=Llama3.1-8B-Instruct2026.05 | 71.03 | 66.55 | 47.24 | 22.76 | 16.55 | 54.48 | — | |
| Reward ModelBase Model=Llama3.1-8B-Instruct2026.05 | 81.38 | 47.59 | 25.17 | 44.14 | 24.48 | 56.9 | — | |
| SQSD(Beaver)Base Model=Qwen3-8B2026.05 | 82.76 | 41.38 | 27.59 | 17.93 | 11.03 | 71.73 | — | |
| SQSD(Aegis)Base Model=Llama3.1-8B-Instruct2026.05 | 82.76 | 26.55 | 52.41 | 25.86 | 10.69 | 72.07 | — | |
| SQSD(Beaver)Base Model=Llama3.1-8B-Instruct2026.05 | 90 | 61.38 | 42.07 | 8.62 | 2.76 | 87.24 | — |