Safety Moderation on WILDJAILBREAK (val)
0.7ASRWILDGUARD
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| WILDGUARDCompletion Model=Tulu2+WJ2024.06 | 0.7 | 1.7 | |
| Aegis-Guard-DCompletion Model=Tulu2+WJ2024.06 | 0.9 | 16.8 | |
| WILDGUARDTarget Model=Llama3-8B-Inst2024.06 | 1.1 | 8 | |
| Aegis-Guard-PCompletion Model=Tulu2+WJ2024.06 | 1.7 | 4 | |
| Llama-Guard2Completion Model=Tulu2+WJ2024.06 | 1.9 | 1.6 | |
| MD-JudgeCompletion Model=Tulu2+WJ2024.06 | 1.9 | 10.1 | |
| No FilterCompletion Model=Tulu2+WJ2024.06 | 2.3 | 1.2 | |
| WILDGUARDCompletion Model=Tulu2-dpo2024.06 | 2.4 | 0.4 | |
| Aegis-Guard-DTarget Model=Llama3-8B-Inst2024.06 | 6 | 20.8 | |
| MD-JudgeTarget Model=Llama3-8B-Inst2024.06 | 10.1 | 10.8 | |
| Aegis-Guard-PTarget Model=Llama3-8B-Inst2024.06 | 12.2 | 10 | |
| Aegis-Guard-DCompletion Model=Tulu2-dpo2024.06 | 12.4 | 16 | |
| Llama-Guard2Target Model=Llama3-8B-Inst2024.06 | 14.2 | 8.4 | |
| NoneTarget Model=Llama3-8B-Inst2024.06 | 17.4 | 7.6 | |
| MD-JudgeCompletion Model=Tulu2-dpo2024.06 | 25.7 | 4.4 | |
| Aegis-Guard-PCompletion Model=Tulu2-dpo2024.06 | 32.7 | 3.6 | |
| Llama-Guard2Completion Model=Tulu2-dpo2024.06 | 53.1 | 0.8 | |
| No FilterCompletion Model=Tulu2-dpo2024.06 | 79.8 | 0 |