Long-context retrieval on NIAH-Single-3 (256K)
76.4Accuracy+QK-Restore
Evaluation Results
| Method | Links | |
|---|---|---|
| +QK-RestoreModel=HypeNet-5B2026.06 | 76.4 | |
| Pre-trainModel=HypeNet-5B2026.06 | 75.2 | |
| +SFTModel=HypeNet-5B2026.06 | 65.4 | |
| Pre-trainModel=HypeNet-9B2026.06 | 52 | |
| +QK-RestoreModel=HypeNet-9B2026.06 | 42.6 | |
| Pre-trainModel=Jet-Nemotron-2B2026.06 | 40.6 | |
| +QK-RestoreModel=Jet-Nemotron-2B2026.06 | 34 | |
| +QK-RestoreModel=HypeNet-2B2026.06 | 30.2 | |
| +SFTModel=Jet-Nemotron-2B2026.06 | 27.8 | |
| Pre-trainModel=HypeNet-2B2026.06 | 23.4 | |
| +SFTModel=HypeNet-9B2026.06 | 22.8 | |
| +SFTModel=HypeNet-2B2026.06 | 10.6 |