Safety Evaluation on HarmBench translated (test)
11Success Rate (EN)Vanilla Model
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Vanilla ModelBase Model=Llama-3.1-8B-Instruct2026.04 | 11 | 16 | 48 | 27 | 58 | 65 | 28.4 | |
| Vanilla ModelBase Model=Qwen-2.5-7B-Instruct2026.04 | 9 | 8 | 19 | 17 | 86 | 52 | 25.1 | |
| CPOBase Model=Qwen-2.5-7B-Instruct2026.04 | 4 | 0 | 13 | 9 | 79 | 38 | 17.5 | |
| CPOBase Model=Llama-3.1-8B-Instruct2026.04 | 3 | 2 | 7 | 3 | 44 | 31 | 10.6 | |
| MPOBase Model=Qwen-2.5-7B-Instruct2026.04 | 3 | 2 | 10 | 6 | 72 | 32 | 14.7 | |
| DPOBase Model=Llama-3.1-8B-Instruct2026.04 | 2 | 7 | 19 | 7 | 29 | 24 | 10.9 | |
| MPOBase Model=Llama-3.1-8B-Instruct2026.04 | 1 | 1 | 10 | 2 | 31 | 19 | 7.6 | |
| LASABase Model=Llama-3.1-8B-Instruct2026.04 | 1 | 0 | 0 | 0 | 16 | 17 | 3.9 | |
| SFTBase Model=Qwen-2.5-7B-Instruct2026.04 | 1 | 0 | 4 | 2 | 67 | 16 | 10.3 | |
| ORPOBase Model=Qwen-2.5-7B-Instruct2026.04 | 1 | 0 | 0 | 0 | 56 | 14 | 7.5 | |
| LASABase Model=Qwen-2.5-7B-Instruct2026.04 | 1 | 0 | 0 | 4 | 25 | 16 | 5.6 | |
| SFTBase Model=Llama-3.1-8B-Instruct2026.04 | 0 | 2 | 6 | 4 | 45 | 29 | 9.7 | |
| KTOBase Model=Llama-3.1-8B-Instruct2026.04 | 0 | 1 | 3 | 2 | 25 | 15 | 5.4 | |
| ORPOBase Model=Llama-3.1-8B-Instruct2026.04 | 0 | 1 | 2.02 | 1.01 | 23 | 15 | 4.3 | |
| DPOBase Model=Qwen-2.5-7B-Instruct2026.04 | 0 | 1 | 8 | 7 | 70 | 33 | 14.5 | |
| KTOBase Model=Qwen-2.5-7B-Instruct2026.04 | 0 | 0 | 7 | 5 | 73 | 28 | 13.5 |