Prohibited Content Detection on ALERT
15ASRICD
Evaluation results
| Method | Links | |
|---|---|---|
| ICDModel=Qwen2.5-7B-Instruct2026.01 | 15 | |
| PPLModel=Qwen2.5-7B-Instruct2026.01 | 14 | |
| Self-ReminderModel=Qwen2.5-7B-Instruct2026.01 | 14 | |
| No DefenseModel=Qwen2.5-7B-Instruct2026.01 | 13.4 | |
| Self-ExaminationModel=Qwen2.5-7B-Instruct2026.01 | 12.8 | |
| PPLModel=Llama-3-8B-Instruct2026.01 | 4.8 | |
| No DefenseModel=Llama-3-8B-Instruct2026.01 | 4.2 | |
| Self-ExaminationModel=Llama-3-8B-Instruct2026.01 | 3.8 | |
| Self-ReminderModel=Llama-3-8B-Instruct2026.01 | 1.4 | |
| SafeDecodingModel=Qwen2.5-7B-Instruct2026.01 | 1 | |
| ICDModel=Llama-3-8B-Instruct2026.01 | 0.8 | |
| SafeDecodingModel=Llama-3-8B-Instruct2026.01 | 0.4 | |
| DPO-HBackbone=Llama3-8B-Instruct2025.12 | 0.2768 | |
| SafeRLHFBackbone=Qwen2.5-7B-Instruct2025.12 | 0.1524 | |
| SafeRLHFBackbone=Llama3-8B-Instruct2025.12 | 0.1366 | |
| DPO-SBackbone=Qwen2.5-7B-Instruct2025.12 | 0.0961 | |
| DPO-MixBackbone=Qwen2.5-7B-Instruct2025.12 | 0.0922 | |
| PeCANBackbone=Qwen2.5-7B-Instruct2025.12 | 0.0865 | |
| MoCANBackbone=Qwen2.5-7B-Instruct2025.12 | 0.0864 | |
| Qwen2.5-7B-InstructBackbone=Qwen2.5-7B-Instruct2025.12 | 0.0862 | |
| DPO-HBackbone=Qwen2.5-7B-Instruct2025.12 | 0.0827 | |
| BFPOBackbone=Llama3-8B-Instruct2025.12 | 0.0722 | |
| BFPOBackbone=Qwen2.5-7B-Instruct2025.12 | 0.0583 | |
| DPO-MixBackbone=Llama3-8B-Instruct2025.12 | 0.0521 | |
| MoCANBackbone=Llama3-8B-Instruct2025.12 | 0.046 | |
| Llama3-8B-InstructBackbone=Llama3-8B-Instruct2025.12 | 0.0381 | |
| PeCANBackbone=Llama3-8B-Instruct2025.12 | 0.0346 | |
| W-DOORBackbone=Qwen2.5-7B-Instruct2025.12 | 0.034 | |
| DPO-SBackbone=Llama3-8B-Instruct2025.12 | 0.0326 | |
| W-DOORBackbone=Llama3-8B-Instruct2025.12 | 0.0186 | |
| NSPOBackbone=Llama3-8B-Instruct2025.12 | 0.01 | |
| NSPOBackbone=Qwen2.5-7B-Instruct2025.12 | 0.0081 | |
| SafeThinkerModel=Llama-3-8B-Instruct2026.01 | 0 | |
| SafeThinkerModel=Qwen2.5-7B-Instruct2026.01 | 0 |