Long-context retrieval on NIAH-Single-3 64K context
97Accuracy+QK-Restore
Evaluation results
| Method | Links | |
|---|---|---|
| +QK-RestoreModel=HypeNet-9B2026.06 | 97 | |
| +SFTModel=HypeNet-5B2026.06 | 96.6 | |
| +QK-RestoreModel=HypeNet-5B2026.06 | 96.4 | |
| Pre-trainModel=HypeNet-9B2026.06 | 95 | |
| Pre-trainModel=HypeNet-5B2026.06 | 94.2 | |
| +SFTModel=HypeNet-9B2026.06 | 90 | |
| +QK-RestoreModel=HypeNet-2B2026.06 | 52.8 | |
| Pre-trainModel=Jet-Nemotron-2B2026.06 | 48 | |
| +QK-RestoreModel=Jet-Nemotron-2B2026.06 | 39 | |
| Pre-trainModel=HypeNet-2B2026.06 | 34.8 | |
| +SFTModel=Jet-Nemotron-2B2026.06 | 33.6 | |
| +SFTModel=HypeNet-2B2026.06 | 32.6 |