Multilingual Safety on MultiJail Out-of-Distribution Languages (test)
0.63Safety Violation Rate (KO)DPO+MLC
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| DPO+MLCBackbone=Qwen 2.5-7B-Instruct2026.02 | 0.63 | 0.63 | 0.32 | 0.95 | 0 | 0.51 | |
| DPO+MLCBackbone=Gemma-2-9B-it2026.02 | 0.63 | 0.32 | 0 | 0.32 | 0.63 | 0.38 | |
| MPOBackbone=Qwen 2.5-7B-Instruct2026.02 | 2.22 | 4.13 | 2.86 | 5.71 | 3.17 | 3.62 | |
| DPOBackbone=Qwen 2.5-7B-Instruct2026.02 | 3.17 | 3.49 | 4.13 | 5.4 | 1.9 | 3.62 | |
| DPOBackbone=Gemma-2-9B-it2026.02 | 3.17 | 0.95 | 5.71 | 1.27 | 2.86 | 2.79 | |
| MPOBackbone=Gemma-2-9B-it2026.02 | 3.49 | 1.9 | 4.13 | 1.27 | 1.59 | 2.48 | |
| RawBackbone=Gemma-2-9B-it2026.02 | 5.08 | 3.49 | 6.67 | 2.22 | 3.17 | 4.13 | |
| RawBackbone=Qwen 2.5-7B-Instruct2026.02 | 5.71 | 9.52 | 10.16 | 8.25 | 6.67 | 8.06 | |
| SDRRLBackbone=Qwen 2.5-7B-Instruct2026.02 | 7.3 | 9.52 | 9.2 | 8.57 | 7.3 | 8.38 | |
| SDRRLBackbone=Gemma-2-9B-it2026.02 | 22.86 | 32.06 | 20.63 | 34.6 | 34.29 | 28.89 |