Instruction Following on Alpaca-Eval (test)
66.85Length-Controlled WinrateNo Defense
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| No DefenseBackbone=Vicuna2026.02 | 66.85 | 66.83 | 1.64 | 805 | 1,108 | |
| NGSDBackbone=Vicuna2026.02 | 65.17 | 64.18 | 1.69 | 805 | 1,015 | |
| No DefenseBackbone=Llama3-8b2026.02 | 63.49 | 64.63 | 1.69 | 805 | 1,616 | |
| NGSDBackbone=Llama3-8b2026.02 | 56.79 | 53.73 | 1.76 | 805 | 1,366 | |
| SSDBackbone=Vicuna2026.02 | 55.87 | 57.89 | 1.74 | 805 | 1,155 | |
| SSDBackbone=Llama3-8b2026.02 | 37 | 37.02 | 1.7 | 805 | 1,378 |