Text Classification on GLUE (SST2, QNLI subset)
95.88SST-2 AccuracyG-Loss (RoBERTa-large)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| G-Loss (RoBERTa-large)Backbone=RoBERTa-large, Loss=GLoss-O+CE2026.04 | 95.88 | 94.27 | |
| RoBERTa-largeLoss=SCL+CE2026.04 | 95.83 | 94.01 | |
| RoBERTa-largeLoss=CE2026.04 | 94.84 | 93.71 | |
| G-Loss (BERT-base-uncased)Backbone=BERT-base-uncased, Loss=GLoss-O+CE2026.04 | 93.32 | 89.62 | |
| BERT-base-uncasedLoss=SCL+CE2026.04 | 92.1 | 90.06 | |
| BERT-base-uncasedLoss=CE2026.04 | 91.2 | 89.08 | |
| G-Loss (DistilBERT-base)Backbone=DistilBERT-base, Loss=GLoss-O+CE2026.04 | 91.06 | 88.34 | |
| DistilBERT-baseLoss=SCL+CE2026.04 | 90.36 | 87.18 | |
| DistilBERT-baseLoss=CE2026.04 | 90.14 | 87.86 |