Preference Alignment on AlpacaEval weighted gpt4 turbo 2.0
46.11Win RateGANPO (SimPO)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| GANPO (SimPO)Backbone=Llama3-8B-Instruct, Disc.=Transformer2026.01 | 46.11 | 50.48 | 1,834 | |
| SimPOBackbone=Llama3-8B-Instruct2026.01 | 44.09 | 48.31 | 1,836 | |
| GANPO (DPO)Backbone=Llama3-8B-Instruct, Disc.=Transformer2026.01 | 35.23 | 33.87 | 2,043 | |
| DPOBackbone=Llama3-8B-Instruct2026.01 | 33.9 | 32.34 | 2,041 | |
| GANPO (SimPO)Backbone=Gemma2-2B-it, Disc.=Transformer2026.01 | 31.37 | 36.74 | 1,745 | |
| SimPOBackbone=Gemma2-2B-it2026.01 | 30.66 | 36.03 | 1,740 | |
| GANPO (DPO)Backbone=Gemma2-2B-it, Disc.=Transformer2026.01 | 24.17 | 29.69 | 1,664 | |
| DPOBackbone=Gemma2-2B-it2026.01 | 22.76 | 27.79 | 1,668 |