Natural Language Understanding on GLUE (SST-2, QQP, QNLI, MNLI Subset)
91.17SST-2 AccuracyDP-FedAdamW
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DP-FedAdamWBackbone=RoBERTa-Base, Data heterogeneity (alpha)=0.8, Evaluation protocol=fine-tuning2026.02 | 91.17 | 83.34 | 86.33 | 78.68 | |
| DP-LocalAdamWBackbone=RoBERTa-Base, Data heterogeneity (alpha)=0.8, Evaluation protocol=fine-tuning2026.02 | 91.04 | 82.33 | 86.07 | 75.2 | |
| DP-FedSAMBackbone=RoBERTa-Base, Data heterogeneity (alpha)=0.8, Evaluation protocol=fine-tuning2026.02 | 70.95 | 75.66 | 70.54 | 56.88 | |
| DP-FedAvg-LSBackbone=RoBERTa-Base, Data heterogeneity (alpha)=0.8, Evaluation protocol=fine-tuning2026.02 | 68.38 | 72.59 | 69.22 | 50.46 | |
| DP-SCAFFOLDBackbone=RoBERTa-Base, Data heterogeneity (alpha)=0.8, Evaluation protocol=fine-tuning2026.02 | 52.35 | 66.87 | 52.62 | 38.74 | |
| DP-FedAvgBackbone=RoBERTa-Base, Data heterogeneity (alpha)=0.8, Evaluation protocol=fine-tuning2026.02 | 49.08 | 63.18 | 49.46 | 35.45 |