Preference Optimization on Roleplay 1500 users
90.9WinrateOracle
Evaluation Results
| Method | Links | |
|---|---|---|
| OracleOracle status=true, Few-shot protocol=Prompting2025.02 | 90.9 | |
| FSPO + RATOptimization algorithm=IPO, Inclusion of RAT=true2025.02 | 90.3 | |
| FSPOOptimization algorithm=IPO2025.02 | 82.6 | |
| FSPOOptimization algorithm=DPO2025.02 | 81.3 | |
| IPOOptimization algorithm=IPO2025.02 | 72.4 | |
| VPLOptimization algorithm=VPL2025.02 | 67.3 | |
| Few-shot PromptingFew-shot protocol=Prompting2025.02 | 63.2 | |
| Few-shot Pref-FTFew-shot protocol=Pref-FT, Optimization algorithm=GPO2025.02 | 62.8 | |
| RICOptimization algorithm=RIC2025.02 | 53.3 | |
| Llama 3.2 3B Instruct2025.02 | 50 |