Instruction Following on AlpacaEval (OOD)
24.9KL Div (α=1)GEB-arctanh(π − 1)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| GEB-arctanh(π − 1)Backbone=LLaMA-3-8B-SFT, Evaluation Protocol=Out-of-domain2025.09 | 24.9 | 25.96 | 19.62 | 23.49 | |
| f-DPOBackbone=LLaMA-3-8B-SFT, Evaluation Protocol=Out-of-domain2025.09 | 25.72 | 24.73 | 17.8 | 22.75 | |
| FEBBackbone=LLaMA-3-8B-SFT, Evaluation Protocol=Out-of-domain2025.09 | 25.72 | 23.75 | 19.62 | 23.03 | |
| GEB-1/πBackbone=LLaMA-3-8B-SFT, Evaluation Protocol=Out-of-domain2025.09 | 26.1 | 25.28 | 19.8 | 23.73 | |
| GEB-πBackbone=LLaMA-3-8B-SFT, Evaluation Protocol=Out-of-domain2025.09 | 28.27 | 25.87 | 20.05 | 24.73 |