Generative Performance on Ultrafeedback 61.1k (test)
69.8Win RateFA-DPO
Evaluation Results
| Method | Links | |
|---|---|---|
| FA-DPOFlipping ratio (η)=30%, Model Backbone=Pythia-1B2025.11 | 69.8 | |
| FA-DPOFlipping ratio (η)=40%, Model Backbone=Pythia-1B2025.11 | 69.8 | |
| FA-DPOFlipping ratio (η)=10%, Model Backbone=Pythia-1B2025.11 | 68.45 | |
| cDPOFlipping ratio (η)=10%, Model Backbone=Pythia-1B2025.11 | 67.8 | |
| ROPOFlipping ratio (η)=40%, Model Backbone=Pythia-1B2025.11 | 67.25 | |
| cDPOFlipping ratio (η)=40%, Model Backbone=Pythia-1B2025.11 | 67.05 | |
| DPOFlipping ratio (η)=0%, Model Backbone=Pythia-1B2025.11 | 67 | |
| FA-DPOFlipping ratio (η)=20%, Model Backbone=Pythia-1B2025.11 | 66.9 | |
| FA-DPOFlipping ratio (η)=0%, Model Backbone=Pythia-1B2025.11 | 66.85 | |
| ROPOFlipping ratio (η)=10%, Model Backbone=Pythia-1B2025.11 | 66.85 | |
| ROPOFlipping ratio (η)=0%, Model Backbone=Pythia-1B2025.11 | 66.8 | |
| cDPOFlipping ratio (η)=0%, Model Backbone=Pythia-1B2025.11 | 66.65 | |
| cDPOFlipping ratio (η)=30%, Model Backbone=Pythia-1B2025.11 | 66.55 | |
| cDPOFlipping ratio (η)=20%, Model Backbone=Pythia-1B2025.11 | 66.05 | |
| rDPOFlipping ratio (η)=0%, Model Backbone=Pythia-1B2025.11 | 65.6 | |
| ROPOFlipping ratio (η)=20%, Model Backbone=Pythia-1B2025.11 | 65.55 | |
| DPOFlipping ratio (η)=30%, Model Backbone=Pythia-1B2025.11 | 64.7 | |
| ROPOFlipping ratio (η)=30%, Model Backbone=Pythia-1B2025.11 | 64.4 | |
| DPOFlipping ratio (η)=40%, Model Backbone=Pythia-1B2025.11 | 64.35 | |
| SIMPOFlipping ratio (η)=0%, Model Backbone=Pythia-1B2025.11 | 62.6 | |
| DPOFlipping ratio (η)=10%, Model Backbone=Pythia-1B2025.11 | 60.1 | |
| rDPOFlipping ratio (η)=30%, Model Backbone=Pythia-1B2025.11 | 57.85 | |
| rDPOFlipping ratio (η)=40%, Model Backbone=Pythia-1B2025.11 | 57.8 | |
| rDPOFlipping ratio (η)=10%, Model Backbone=Pythia-1B2025.11 | 56.65 | |
| SIMPOFlipping ratio (η)=20%, Model Backbone=Pythia-1B2025.11 | 56.5 | |
| DPOFlipping ratio (η)=20%, Model Backbone=Pythia-1B2025.11 | 56.35 | |
| SIMPOFlipping ratio (η)=40%, Model Backbone=Pythia-1B2025.11 | 55.8 | |
| SIMPOFlipping ratio (η)=30%, Model Backbone=Pythia-1B2025.11 | 54.9 | |
| rDPOFlipping ratio (η)=20%, Model Backbone=Pythia-1B2025.11 | 54.85 | |
| SIMPOFlipping ratio (η)=10%, Model Backbone=Pythia-1B2025.11 | 51.2 |