Preference Alignment Evaluation on PRISM (test)
0.331BT Score (Mean)Hard Panel
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Hard PanelBase Model=Llama-3.2-3B, Bootstrap size (n)=10002026.02 | 0.331 | — | 0.279 | — | |
| Hard PanelBase Model=Llama-3.2-1B, Bootstrap samples=10002026.02 | 0.267 | — | 0.215 | — | |
| US-RepBase Model=Llama-3.2-1B, Bootstrap samples=10002026.02 | 0.146 | — | 0.132 | — | |
| US-RepBase Model=Llama-3.2-3B, Bootstrap size (n)=10002026.02 | 0.135 | — | 0.106 | — | |
| Soft PanelBase Model=Llama-3.2-1B, Bootstrap samples=10002026.02 | -0.106 | — | -0.017 | — | |
| Soft PanelBase Model=Llama-3.2-3B, Bootstrap size (n)=10002026.02 | -0.108 | — | -0.026 | — | |
| Full PRISMBase Model=Llama-3.2-1B, Bootstrap samples=10002026.02 | -0.127 | — | -0.047 | — | |
| Full PRISMBase Model=Llama-3.2-3B, Bootstrap size (n)=10002026.02 | -0.133 | — | -0.057 | — | |
| BaseBase Model=Llama-3.2-1B, Bootstrap samples=10002026.02 | -0.18 | — | -0.283 | — | |
| BaseBase Model=Llama-3.2-3B, Bootstrap size (n)=10002026.02 | -0.225 | — | -0.302 | — |