3-shot Evaluation on ARC, BoolQ, HSwag, LAMBADA, OBQA, PIQA, and Wino Suite
14.18Lambada PerplexityQK-Normed MLA
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| QK-Normed MLAshots=3-shot, checkpoints=400M/100B2026.06 | 14.18 | 24.66 | 56.69 | 62.2 | 38.94 | 43.92 | 19.6 | 70.89 | 53.75 | 46.33 | |
| QK-Clipshots=3-shot, checkpoints=400M/100B2026.06 | 16.28 | 22.61 | 53.07 | 59.51 | 38.32 | 42.4 | 18.2 | 70.67 | 53.2 | 44.75 |