Long-context retrieval on NIAH-Single-3 128K context
90.6Accuracy+SFT
Evaluation Results
| Method | Links | |
|---|---|---|
| +SFTModel=HypeNet-5B2026.06 | 90.6 | |
| +QK-RestoreModel=HypeNet-5B2026.06 | 90.2 | |
| Pre-trainModel=HypeNet-5B2026.06 | 87.2 | |
| Pre-trainModel=HypeNet-9B2026.06 | 83.8 | |
| Pre-trainModel=Jet-Nemotron-2B2026.06 | 73.6 | |
| +SFTModel=Jet-Nemotron-2B2026.06 | 67.2 | |
| +QK-RestoreModel=HypeNet-9B2026.06 | 66.4 | |
| +QK-RestoreModel=Jet-Nemotron-2B2026.06 | 64.4 | |
| +SFTModel=HypeNet-9B2026.06 | 47.8 | |
| +QK-RestoreModel=HypeNet-2B2026.06 | 30.8 | |
| Pre-trainModel=HypeNet-2B2026.06 | 30.4 | |
| +SFTModel=HypeNet-2B2026.06 | 7.6 |