Long-context retrieval and aggregation on RULER 8K context length
95.4NIAH-MVFOCUSFT
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| FOCUSFT2026.05 | 95.4 | 88.8 | 97.5 | |
| Pretrained2026.05 | 95.1 | 83.2 | 96.5 | |
| Standard SFT2026.05 | 94.6 | 85.4 | 97.4 |
| Method | Links | |||
|---|---|---|---|---|
| FOCUSFT2026.05 | 95.4 | 88.8 | 97.5 | |
| Pretrained2026.05 | 95.1 | 83.2 | 96.5 | |
| Standard SFT2026.05 | 94.6 | 85.4 | 97.4 |