Co-reference Resolution on synthetic dataset
81AccuracyOracle
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| OracleModel size (Llama)=3B2025.03 | 81 | 95.38 | |
| TokenButlerModel size (Llama)=3B2025.03 | 80.2 | 95.23 | |
| OracleModel size (Llama)=8B2025.03 | 77 | 93.47 | |
| TokenButlerModel size (Llama)=8B2025.03 | 76.17 | 92.59 | |
| OracleModel size (Llama)=1B2025.03 | 49 | 84.32 | |
| TokenButlerModel size (Llama)=1B2025.03 | 48.94 | 84.02 | |
| Token EvictionModel size (Llama)=3B2025.03 | 10 | 51.97 | |
| Page-BasedModel size (Llama)=3B2025.03 | 6 | 57.82 | |
| Token EvictionModel size (Llama)=8B2025.03 | 3 | 37.5 | |
| Token EvictionModel size (Llama)=1B2025.03 | 1 | 32.5 | |
| Page-BasedModel size (Llama)=1B2025.03 | 0 | 19.78 | |
| Page-BasedModel size (Llama)=8B2025.03 | 0 | 46.98 |