Natural Language Understanding on GLUE (test) (All 9 Tasks + Avg)
91.55MNLI ScoreRiemannion (SGD)
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Riemannion (SGD)Backbone=Mistral-7B, Quantization=4-bit NF4, LoRA rank=16, Optimizer momentum=None2026.05 | 91.55 | 96.9 | 82.6 | 68.9 | 94.93 | 87.77 | 82.67 | 77.99 | 60.56 | 82.65 | |
| Scaled GDBackbone=Mistral-7B, Quantization=4-bit NF4, LoRA rank=16, Optimizer momentum=None2026.05 | 90.21 | 96.9 | 81.62 | 68.17 | 94.4 | 91.15 | 54.15 | 90.31 | 56.34 | 80.36 | |
| Factor-wise MuonBackbone=Mistral-7B, Quantization=4-bit NF4, LoRA rank=16, Optimizer momentum=None2026.05 | 89.77 | 97.25 | 82.35 | 69.32 | 94.47 | 88.42 | 84.12 | 84.39 | 45.07 | 81.68 | |
| iMuonBackbone=Mistral-7B, Quantization=4-bit NF4, LoRA rank=16, Optimizer momentum=None2026.05 | 89.29 | 97.02 | 85.54 | 70.21 | 94.6 | 87.86 | 88.09 | 88.99 | 74.65 | 86.25 | |
| SGDBackbone=Mistral-7B, Quantization=4-bit NF4, LoRA rank=16, Optimizer momentum=None2026.05 | 88.15 | 96.1 | 70.1 | 55.89 | 94.22 | 88.59 | 50.9 | 47.64 | 49.3 | 71.21 |