Safety Evaluation on Alpaca
0HRRBefore Fine-tuning
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Before Fine-tuningBackbone=Llama-2-chat-7B2026.04 | 0 | — | — | |
| Ours (I)Backbone=Llama-2-chat-7B, Rank=322026.04 | 0 | — | — | |
| Ours (cannot)Backbone=Llama-2-chat-7B, Rank=322026.04 | 0.6 | — | — | |
| Before Fine-tuningBackbone=Llama-3.1-Instruct-8B2026.04 | 1.4 | — | — | |
| SaLoRABackbone=Llama-3.1-Instruct-8B, Rank=322026.04 | 1.4 | — | — | |
| Ours (I)Backbone=Llama-3.1-Instruct-8B, Rank=322026.04 | 1.5 | — | — | |
| Ours (cannot)Backbone=Llama-3.1-Instruct-8B, Rank=322026.04 | 1.5 | — | — | |
| SaLoRABackbone=Llama-3.1-Instruct-8B, Rank=162026.04 | 2.9 | — | — | |
| Ours (I)Backbone=Llama-2-chat-7B, Rank=162026.04 | 2.9 | — | — | |
| Ours (I)Backbone=Llama-3.1-Instruct-8B, Rank=162026.04 | 3.1 | — | — | |
| Ours (cannot)Backbone=Llama-2-chat-7B, Rank=162026.04 | 3.4 | — | — | |
| SaLoRABackbone=Llama-2-chat-7B, Rank=162026.04 | 3.5 | — | — | |
| Ours (cannot)Backbone=Llama-3.1-Instruct-8B, Rank=162026.04 | 3.5 | — | — | |
| SaLoRABackbone=Llama-2-chat-7B, Rank=322026.04 | 4.4 | — | — | |
| LORA w. IABackbone=Llama-3.1-Instruct-8B, Rank=322026.04 | 5.8 | — | — | |
| Safe LoRABackbone=Llama-3.1-Instruct-8B, Rank=322026.04 | 6.7 | — | — | |
| LORA w. IABackbone=Llama-3.1-Instruct-8B, Rank=162026.04 | 7.7 | — | — | |
| Safe LoRABackbone=Llama-3.1-Instruct-8B, Rank=162026.04 | 8.5 | — | — | |
| DORABackbone=Llama-3.1-Instruct-8B, Rank=322026.04 | 9.4 | — | — | |
| DORABackbone=Llama-3.1-Instruct-8B, Rank=162026.04 | 10.1 | — | — | |
| LORA w. IABackbone=Llama-2-chat-7B, Rank=162026.04 | 13.5 | — | — | |
| LORABackbone=Llama-3.1-Instruct-8B, Rank=162026.04 | 13.8 | — | — | |
| PiSSABackbone=Llama-3.1-Instruct-8B, Rank=162026.04 | 13.8 | — | — | |
| LORABackbone=Llama-3.1-Instruct-8B, Rank=322026.04 | 14.5 | — | — | |
| PiSSABackbone=Llama-3.1-Instruct-8B, Rank=322026.04 | 14.5 | — | — | |
| Safe LoRABackbone=Llama-2-chat-7B, Rank=322026.04 | 14.5 | — | — | |
| Safe LoRABackbone=Llama-2-chat-7B, Rank=162026.04 | 15.7 | — | — | |
| LORA w. VacBackbone=Llama-2-chat-7B, Rank=162026.04 | 20.2 | — | — | |
| LORABackbone=Llama-2-chat-7B, Rank=162026.04 | 23.7 | — | — | |
| DORABackbone=Llama-2-chat-7B, Rank=162026.04 | 23.7 | — | — | |
| LORA w. IABackbone=Llama-2-chat-7B, Rank=322026.04 | 23.7 | — | — | |
| DORABackbone=Llama-2-chat-7B, Rank=322026.04 | 25.3 | — | — | |
| LORA w. VacBackbone=Llama-2-chat-7B, Rank=322026.04 | 25.3 | — | — | |
| LORABackbone=Llama-2-chat-7B, Rank=322026.04 | 31.7 | — | — | |
| PiSSABackbone=Llama-2-chat-7B, Rank=162026.04 | 31.7 | — | — | |
| PiSSABackbone=Llama-2-chat-7B, Rank=322026.04 | 35.7 | — | — | |
| LORA w. VacBackbone=Llama-3.1-Instruct-8B, Rank=322026.04 | 38.3 | — | — | |
| LORA w. VacBackbone=Llama-3.1-Instruct-8B, Rank=162026.04 | 41.1 | — | — | |
| Goal PrioritizationTarget Model=Vicuna-7B2026.04 | — | 13 | 0.1 | |
| Llama Guard 3Target Model=Vicuna-7B2026.04 | — | 0 | 38.5 | |
| No DefenseTarget Model=Vicuna-7B2026.04 | — | 0 | 0 | |
| Qwen3GuardTarget Model=Vicuna-7B2026.04 | — | 0 | 42.8 | |
| Self-GuardTarget Model=Vicuna-7B2026.04 | — | 7.5 | 24.6 | |
| TrajGuardTarget Model=Vicuna-7B2026.04 | — | 0 | 5.2 |