Preference Alignment on EduFeedback-Alternate (test)
67.78Pairwise Win Rate (excl. ties)DPO
Evaluation Results
| Method | Links | |
|---|---|---|
| DPOBackbone=Llama-3.2-3B, Evaluation Protocol=ArenaHard Custom Pairwise WR ex Ties2026.05 | 67.78 | |
| COALABackbone=Llama-3.2-3B, Evaluation Protocol=ArenaHard Custom Pairwise WR ex Ties2026.05 | 67.5 | |
| ORPOBackbone=Llama-3.2-3B, Evaluation Protocol=ArenaHard Custom Pairwise WR ex Ties2026.05 | 56.25 | |
| SFTBackbone=Llama-3.2-3B, Evaluation Protocol=ArenaHard Custom Pairwise WR ex Ties2026.05 | 27.78 | |
| SimPOBackbone=Llama-3.2-3B, Evaluation Protocol=ArenaHard Custom Pairwise WR ex Ties2026.05 | 10 |