Needle-in-a-haystack retrieval on RULER S-NIAH-1 OOD
100Success Rate (4K Context)ASEntmax
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ASEntmaxPositional encoding=NAPE, Training context length=20482025.06 | 100 | 99.8 | 97.4 | |
| SSMaxPositional encoding=NAPE, Training context length=20482025.06 | 99.2 | 92 | 75.2 | |
| SoftmaxPositional encoding=NAPE, Training context length=20482025.06 | 94.2 | 11.4 | 0.8 | |
| EntmaxPositional encoding=NAPE, Training context length=20482025.06 | 89 | 21.6 | 1.2 |