Format Debiasing on UltraFeedback Format-Biased (test)
89Win-Rate (Bold)Bradley-Terry (BT)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Bradley-Terry (BT)training=standard2025.12 | 89 | 92.5 | |
| DIR2025.12 | 51.2 | 52 | |
| Format Decoupling (FD)2025.12 | 50.5 | 53 | |
| Bradley-Terry (BT†)training=filtered (removing format-biased examples)2025.12 | 49 | 52.5 |