Natural Language Understanding on GLUE (val) - Subset (CoLA, MRPC, QNLI, QQP, RTE, SST-2, STS-B)
85.15Average ScoreMGUP-AdamW
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| MGUP-AdamWBackbone=RoBERTa-base, Evaluation Protocol=Fine-tuning2026.06 | 85.15 | 75.81 | 90.44 | 90.54 | 59.83 | 94.95 | 93.08 | 91.43 | |
| AdamWBackbone=RoBERTa-base, Evaluation Protocol=Fine-tuning2026.06 | 84.74 | 72.93 | 90.44 | 90.55 | 60.32 | 94.84 | 92.79 | 91.34 | |
| MGUP-LionBackbone=RoBERTa-base, Evaluation Protocol=Fine-tuning2026.06 | 84.18 | 71.12 | 88.24 | 90.07 | 61.23 | 94.27 | 93.04 | 91.33 | |
| MGUP-MuonBackbone=RoBERTa-base, Evaluation Protocol=Fine-tuning2026.06 | 84.17 | 70.4 | 88.24 | 89.84 | 61.07 | 94.61 | 93.24 | 91.78 | |
| C-MuonBackbone=RoBERTa-base, Evaluation Protocol=Fine-tuning2026.06 | 83.98 | 70.04 | 88.24 | 90.04 | 59.81 | 94.84 | 93.19 | 91.75 | |
| C-AdamWBackbone=RoBERTa-base, Evaluation Protocol=Fine-tuning2026.06 | 83.69 | 71.12 | 89.22 | 90.25 | 57.29 | 93.92 | 92.62 | 91.39 | |
| LDAdamW(r=8)Backbone=RoBERTa-base, Evaluation Protocol=Fine-tuning2026.06 | 83.58 | 67.58 | 88.32 | 90.03 | 60.6 | 94.49 | 92.82 | 91.23 | |
| LionBackbone=RoBERTa-base, Evaluation Protocol=Fine-tuning2026.06 | 83.39 | 67.15 | 87.5 | 89.39 | 60.57 | 94.84 | 93 | 91.32 | |
| C-LionBackbone=RoBERTa-base, Evaluation Protocol=Fine-tuning2026.06 | 83.23 | 67.87 | 88.73 | 89.58 | 57.78 | 94.5 | 92.81 | 91.41 | |
| GaLore(r=8)Backbone=RoBERTa-base, Evaluation Protocol=Fine-tuning2026.06 | 82.25 | 69.45 | 86.19 | 88.97 | 55.12 | 94.15 | 92.01 | 89.86 | |
| MuonBackbone=RoBERTa-base, Evaluation Protocol=Fine-tuning2026.06 | 81.65 | 64.62 | 81.13 | 87.33 | 59.34 | 94.27 | 93.11 | 91.72 | |
| Adam-miniBackbone=RoBERTa-base, Evaluation Protocol=Fine-tuning2026.06 | 80.44 | 56.32 | 87.01 | 89.49 | 56.32 | 93.35 | 92.02 | 89.58 |