Safety alignment evaluation on Llama-Guard
82.14Harmfulness (%)IPO
Evaluation Results
| Method | Links | |
|---|---|---|
| IPOBackbone=Llama3.1-8B-Base2025.11 | 82.14 | |
| DPOBackbone=Llama3.1-8B-Base2025.11 | 83.06 | |
| KTOBackbone=Llama3.1-8B-Base2025.11 | 83.42 | |
| DPO-RandomBackbone=Llama3.1-8B-Base2025.11 | 83.42 | |
| SimPOBackbone=Llama3.1-8B-Base2025.11 | 84.15 | |
| RTOBackbone=Llama3.1-8B-Base2025.11 | 85.71 | |
| TDPOBackbone=Llama3.1-8B-Base2025.11 | 86.54 | |
| TIS-DPOBackbone=Llama3.1-8B-Base2025.11 | 87.87 | |
| AAOBackbone=Llama3.1-8B-Base2025.11 | 91.29 | |
| RTOBackbone=Mistral-7B-Base2025.11 | 94.36 | |
| TDPOBackbone=Mistral-7B-Base2025.11 | 94.55 | |
| KTOBackbone=Mistral-7B-Base2025.11 | 95.04 | |
| DPOBackbone=Mistral-7B-Base2025.11 | 95.48 | |
| DPO-RandomBackbone=Mistral-7B-Base2025.11 | 95.52 | |
| DPOBackbone=Llama3.1-8B-Instruct2025.11 | 96.12 | |
| IPOBackbone=Mistral-7B-Base2025.11 | 96.34 | |
| KTOBackbone=Llama3.1-8B-Instruct2025.11 | 96.52 | |
| SimPOBackbone=Mistral-7B-Base2025.11 | 96.56 | |
| TIS-DPOBackbone=Mistral-7B-Base2025.11 | 96.74 | |
| RTOBackbone=Llama3.1-8B-Instruct2025.11 | 96.99 | |
| TDPOBackbone=Llama3.1-8B-Instruct2025.11 | 97.04 | |
| DPO-RandomBackbone=Llama3.1-8B-Instruct2025.11 | 97.32 | |
| TIS-DPOBackbone=Mistral-7B-Instruct2025.11 | 97.35 | |
| AAOBackbone=Mistral-7B-Base2025.11 | 97.41 | |
| IPOBackbone=Llama3.1-8B-Instruct2025.11 | 97.42 | |
| TIS-DPOBackbone=Llama3.1-8B-Instruct2025.11 | 97.54 | |
| SimPOBackbone=Llama3.1-8B-Instruct2025.11 | 97.85 | |
| TDPOBackbone=Mistral-7B-Instruct2025.11 | 97.97 | |
| AAOBackbone=Llama3.1-8B-Instruct2025.11 | 98.42 | |
| DPO-RandomBackbone=Mistral-7B-Instruct2025.11 | 98.42 | |
| RTOBackbone=Mistral-7B-Instruct2025.11 | 98.76 | |
| DPOBackbone=Mistral-7B-Instruct2025.11 | 99.14 | |
| KTOBackbone=Mistral-7B-Instruct2025.11 | 99.21 | |
| SimPOBackbone=Mistral-7B-Instruct2025.11 | 99.35 | |
| IPOBackbone=Mistral-7B-Instruct2025.11 | 99.42 | |
| AAOBackbone=Mistral-7B-Instruct2025.11 | 99.67 |