Aggregate Performance on Average (SST2, AGNEWS, GSM8K, AlpacaEval)
35.5HS ScoreSFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFTBackbone=Llama-2-7B2025.06 | 35.5 | 63.1 | |
| Lisa-baseBackbone=Llama-2-7B2025.06 | 28.8 | 57.56 | |
| BEABackbone=Llama-2-7B2025.06 | 17.05 | 62.36 | |
| Safe LoRABackbone=Llama-2-7B2025.06 | 14.5 | 62.19 | |
| SafeInstrBackbone=Llama-2-7B2025.06 | 13.6 | 59.91 | |
| Lisa-alignedBackbone=Llama-2-7B2025.06 | 10.8 | 63.02 | |
| AsFTBackbone=Llama-2-7B2025.06 | 6.9 | 65.63 |