Safety Evaluation on CategoricalHarmfulQA Alpaca fine-tuning (test)
-22.91ASR Delta (S1-S5)Self-Inf-N
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Self-Inf-NBase Model=Qwen3-8B2026.05 | -22.91 | 2.91 | 1.45 | 7.09 | 7.09 | 25.82 | — | — | — | — | — | — | |
| Bi-Anchor(Reps)Base Model=Llama3.1-8B-Instruct2026.05 | -22 | 20 | 17.45 | 32.36 | 20.18 | 42 | — | — | — | — | — | — | |
| Bi-Anchor(Grad)Base Model=Qwen3-8B2026.05 | -3.46 | 2.36 | 4.55 | 26.36 | 8 | 5.82 | — | — | — | — | — | — | |
| Self-Inf-NBackbone=Qwen3-8B2026.05 | -2.21 | — | — | — | — | — | — | 2.9 | 3.32 | 2.45 | 1.21 | 0.69 | |
| Bi-Anchor(Reps)Backbone=Llama3.1-8B-Instruct2026.05 | -1.95 | — | — | — | — | — | — | 0.45 | 0.94 | -0.56 | 0.18 | -1.5 | |
| LARFBase Model=Llama2-7b-Chat2026.05 | -1.63 | 0.73 | 2.18 | 2.36 | 6.73 | 2.36 | — | — | — | — | — | — | |
| Self-Inf-NBase Model=Llama2-7b-Chat2026.05 | -1.09 | 3.64 | 0.55 | 2.73 | 5.82 | 4.73 | — | — | — | — | — | — | |
| Bi-Anchor(Grad)Base Model=Llama2-7b-Chat2026.05 | -0.73 | 5.27 | 3.64 | 1.45 | 1.64 | 6 | — | — | — | — | — | — | |
| Bi-Anchor(Grad)Backbone=Llama2-7b-Chat2026.05 | -0.69 | — | — | — | — | — | — | 4.34 | 4.24 | 4.82 | 4.35 | 3.65 | |
| Bi-Anchor(Grad)Backbone=Llama3.1-8B-Instruct2026.05 | -0.6 | — | — | — | — | — | — | 0.13 | -0.03 | 0.52 | -1.22 | -0.47 | |
| Bi-Anchor(Grad)Backbone=Qwen3-8B2026.05 | -0.22 | — | — | — | — | — | — | 2.92 | 2.9 | 0.7 | 2.26 | 2.7 | |
| LARFBackbone=Llama2-7b-Chat2026.05 | -0.11 | — | — | — | — | — | — | 4.36 | 4.18 | 4.24 | 3.71 | 4.25 | |
| Self-Inf-NBackbone=Llama2-7b-Chat2026.05 | 0.18 | — | — | — | — | — | — | 3.83 | 4.96 | 4.3 | 4.24 | 4.01 | |
| Self-Inf-NBackbone=Llama3.1-8B-Instruct2026.05 | 0.31 | — | — | — | — | — | — | -1.56 | 0.09 | -0.67 | -2.34 | -1.25 | |
| Bi-Anchor(Reps)Backbone=Qwen3-8B2026.05 | 0.66 | — | — | — | — | — | — | 2.28 | 1.78 | 1.18 | 2.48 | 2.94 | |
| LARFBackbone=Qwen3-8B2026.05 | 1.35 | — | — | — | — | — | — | 1.86 | 1.17 | 3.04 | 3.38 | 3.21 | |
| Bi-Anchor(Reps)Backbone=Llama2-7b-Chat2026.05 | 1.94 | — | — | — | — | — | — | 3.26 | 2.61 | 3.97 | 5.22 | 5.2 | |
| Reward ModelBackbone=Llama2-7b-Chat2026.05 | 3.05 | — | — | — | — | — | — | 2.24 | 3.9 | 4.63 | 5.52 | 5.29 | |
| SQSD(Aegis)Backbone=Llama3.1-8B-Instruct2026.05 | 3.72 | — | — | — | — | — | — | -2.23 | -0.09 | 0.18 | 0.4 | 1.49 | |
| SQSD(Aegis)Backbone=Qwen3-8B2026.05 | 4.91 | — | — | — | — | — | — | -1.21 | 1.22 | 2.22 | 2.45 | 3.7 | |
| SQSD(Beaver)Backbone=Llama2-7b-Chat2026.05 | 4.92 | — | — | — | — | — | — | 0.47 | 4.75 | 4.21 | 5.12 | 5.39 | |
| Reward ModelBackbone=Qwen3-8B2026.05 | 5.02 | — | — | — | — | — | — | -2 | 2.42 | 1.99 | 1.93 | 3.02 | |
| Reward ModelBackbone=Llama3.1-8B-Instruct2026.05 | 5.32 | — | — | — | — | — | — | -3.46 | 1.17 | -0.32 | 1.23 | 1.86 | |
| SQSD(Beaver)Backbone=Qwen3-8B2026.05 | 5.53 | — | — | — | — | — | — | -2.1 | 1.17 | 0.91 | 2.58 | 3.43 | |
| SQSD(Aegis)Backbone=Llama2-7b-Chat2026.05 | 5.57 | — | — | — | — | — | — | -0.33 | 4.07 | 4.45 | 4.8 | 5.24 | |
| LARFBackbone=Llama3.1-8B-Instruct2026.05 | 5.82 | — | — | — | — | — | — | -3.64 | -0.32 | -0.69 | 1.74 | 2.18 | |
| Bi-Anchor(Reps)Base Model=Qwen3-8B2026.05 | 6.18 | 8.73 | 15.27 | 14.73 | 5.45 | 2.55 | — | — | — | — | — | — | |
| Bi-Anchor(Grad)Base Model=Llama3.1-8B-Instruct2026.05 | 6.36 | 33.09 | 27.27 | 18.18 | 37.64 | 26.73 | — | — | — | — | — | — | |
| SQSD(Beaver)Backbone=Llama3.1-8B-Instruct2026.05 | 6.4 | — | — | — | — | — | — | -3.05 | -0.48 | 0.13 | 1.02 | 3.35 | |
| Bi-Anchor(Reps)Base Model=Llama2-7b-Chat2026.05 | 9.46 | 10.91 | 14 | 3.27 | 0.73 | 1.45 | — | — | — | — | — | — | |
| LARFBase Model=Qwen3-8B2026.05 | 11.82 | 13.82 | 22.36 | 4.36 | 1.09 | 2 | — | — | — | — | — | — | |
| Reward ModelBase Model=Llama2-7b-Chat2026.05 | 18.19 | 18.55 | 3.82 | 1.45 | 0 | 0.36 | — | — | — | — | — | — | |
| SQSD(Beaver)Base Model=Llama2-7b-Chat2026.05 | 29.82 | 30 | 1.45 | 2 | 0 | 0.18 | — | — | — | — | — | — | |
| SQSD(Aegis)Base Model=Qwen3-8B2026.05 | 38.36 | 40.91 | 14.36 | 7.27 | 4 | 2.55 | — | — | — | — | — | — | |
| Self-Inf-NBase Model=Llama3.1-8B-Instruct2026.05 | 38.36 | 77.27 | 32.55 | 31.82 | 56.36 | 38.91 | — | — | — | — | — | — | |
| SQSD(Aegis)Base Model=Llama2-7b-Chat2026.05 | 38.91 | 39.27 | 3.27 | 1.09 | 0.55 | 0.36 | — | — | — | — | — | — | |
| SQSD(Aegis)Base Model=Llama3.1-8B-Instruct2026.05 | 41.09 | 49.64 | 29.82 | 23.09 | 16.91 | 8.55 | — | — | — | — | — | — | |
| Reward ModelBase Model=Qwen3-8B2026.05 | 43.82 | 50.73 | 6 | 9.09 | 9.82 | 6.91 | — | — | — | — | — | — | |
| SQSD(Beaver)Base Model=Qwen3-8B2026.05 | 47.64 | 50.91 | 19.09 | 18.73 | 7.27 | 3.27 | — | — | — | — | — | — | |
| Reward ModelBase Model=Llama3.1-8B-Instruct2026.05 | 54.36 | 70.18 | 7.45 | 34 | 15.45 | 15.82 | — | — | — | — | — | — | |
| LARFBase Model=Llama3.1-8B-Instruct2026.05 | 59.82 | 64.73 | 35.27 | 44.36 | 12.55 | 4.91 | — | — | — | — | — | — | |
| SQSD(Beaver)Base Model=Llama3.1-8B-Instruct2026.05 | 65.46 | 65.82 | 27.27 | 25.45 | 9.64 | 0.36 | — | — | — | — | — | — |