Function Addition on MMLU Safety Alignment
44.67Change in RA (%)CNT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CNTModel=Hermes-2-Pro-Llama-3-8B, Transfer Rate (TR)=0.000232026.03 | 44.67 | 0.23 | |
| CNTModel=LLaMA-3-8B-SFR-Iterative-DPO, Transfer Rate (TR)=0.000122026.03 | 37 | 0.28 | |
| LoRA SFTModel=Hermes-2-Pro-Llama-3-8B2026.03 | 34.34 | -0.22 | |
| LoRA SFTModel=LLaMA-3-8B-SFR-Iterative-DPO2026.03 | 33.5 | -0.16 | |
| Model SurgeryModel=LLaMA-3-8B-SFR-Iterative-DPO2026.03 | 30.83 | -12.48 | |
| Model SurgeryModel=Hermes-2-Pro-Llama-3-8B2026.03 | 26.67 | -4.5 |