Safety-Utility Trade-off on IFEval and AdvBench
84.62SUT ScoreCSULoRA
Evaluation Results
| Method | Links | |
|---|---|---|
| CSULoRAModel=Gemma-3-4B-it2026.05 | 84.62 | |
| SafeLoRAModel=Gemma-3-4B-it2026.05 | 81.62 | |
| CSULoRAModel=Llama-3.2-3B-Instruct2026.05 | 77.83 | |
| Base modelModel=Gemma-3-4B-it2026.05 | 74.59 | |
| Base modelModel=Llama-3.2-3B-Instruct2026.05 | 71.59 | |
| RESTAModel=Llama-3.2-3B-Instruct2026.05 | 63.02 | |
| RESTAModel=Gemma-3-4B-it2026.05 | 59.42 | |
| SaLoRAModel=Llama-3.2-3B-Instruct2026.05 | 48.48 | |
| SPLoRAModel=Gemma-3-4B-it2026.05 | 43.22 | |
| LoRAModel=Gemma-3-4B-it2026.05 | 43.18 | |
| AlignGuardModel=Gemma-3-4B-it2026.05 | 42.84 | |
| SPLoRAModel=Llama-3.2-3B-Instruct2026.05 | 38.73 | |
| SaLoRAModel=Gemma-3-4B-it2026.05 | 38.57 | |
| LoRAModel=Llama-3.2-3B-Instruct2026.05 | 32.7 | |
| SafeLoRAModel=Llama-3.2-3B-Instruct2026.05 | 30.64 | |
| AlignGuardModel=Llama-3.2-3B-Instruct2026.05 | 23.27 |