Safety Evaluation on HarmBench, TruthfulQA, and BBQ
91.4Safety ScoreBase (no adapt.)
Evaluation results
| Method | Links | |
|---|---|---|
| Base (no adapt.)Backbone=Llama-2-7B-Chat, Adaptation Protocol=No adaptation2026.04 | 91.4 | |
| Base (no adapt.)Backbone=Mistral-7B-Instruct2026.04 | 88.7 | |
| SafeAnchorBackbone=Llama-2-7B-Chat, Adaptation Protocol=Sequential (Medical → Legal → Code)2026.04 | 85.2 | |
| SafeAnchorBackbone=Mistral-7B-Instruct2026.04 | 82.6 | |
| SafeGrad + LoRABackbone=Llama-2-7B-Chat, Adaptation Protocol=Sequential (Medical → Legal → Code)2026.04 | 67.4 | |
| Safety InterleavingBackbone=Llama-2-7B-Chat, Adaptation Protocol=Sequential (Medical → Legal → Code)2026.04 | 64.8 | |
| SafeGrad + LoRABackbone=Mistral-7B-Instruct2026.04 | 63.8 | |
| Safe LoRABackbone=Llama-2-7B-Chat, Adaptation Protocol=Sequential (Medical → Legal → Code)2026.04 | 61.3 | |
| Safety InterleavingBackbone=Mistral-7B-Instruct2026.04 | 61.2 | |
| Vaccine + LoRABackbone=Llama-2-7B-Chat, Adaptation Protocol=Sequential (Medical → Legal → Code)2026.04 | 58.9 | |
| Safe LoRABackbone=Mistral-7B-Instruct2026.04 | 57.6 | |
| EWC + LoRABackbone=Llama-2-7B-Chat, Adaptation Protocol=Sequential (Medical → Legal → Code)2026.04 | 52.1 | |
| EWC + LoRABackbone=Mistral-7B-Instruct2026.04 | 48.8 | |
| O-LoRABackbone=Llama-2-7B-Chat, Adaptation Protocol=Sequential (Medical → Legal → Code)2026.04 | 48.7 | |
| Standard LoRABackbone=Llama-2-7B-Chat, Adaptation Protocol=Sequential (Medical → Legal → Code)2026.04 | 43.6 | |
| Standard LoRABackbone=Mistral-7B-Instruct2026.04 | 39.2 |