Long-context retrieval on NIAH-Single-2 64K
100Accuracy+SFT
Evaluation Results
| Method | Links | |
|---|---|---|
| +SFTModel=HypeNet-5B2026.06 | 100 | |
| +QK-RestoreModel=HypeNet-5B2026.06 | 100 | |
| +QK-RestoreModel=HypeNet-9B2026.06 | 100 | |
| Pre-trainModel=HypeNet-5B2026.06 | 99.8 | |
| Pre-trainModel=HypeNet-9B2026.06 | 95.8 | |
| Pre-trainModel=HypeNet-2B2026.06 | 94 | |
| +QK-RestoreModel=HypeNet-2B2026.06 | 91.2 | |
| Pre-trainModel=Jet-Nemotron-2B2026.06 | 91 | |
| +SFTModel=HypeNet-2B2026.06 | 81.8 | |
| +SFTModel=HypeNet-9B2026.06 | 64.2 | |
| +QK-RestoreModel=Jet-Nemotron-2B2026.06 | 34 | |
| +SFTModel=Jet-Nemotron-2B2026.06 | 28.4 |