Question Answering on Adversarial SQuAD AddOneSent
81.4EMALUM-RoBERTa-Large
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ALUM-RoBERTa-LargeBackbone=RoBERTa-Large, Training=Adversarial pre-training2020.04 | 81.4 | 75 | |
| RoBERTa-LargeBackbone=RoBERTa-Large2020.04 | 79.3 | 72.9 | |
| InfoBERTTraining protocol=Standard Training, Model backbone=RoBERTa_Large2020.10 | 78.3 | 84.6 | |
| InfoBERTTraining protocol=Adversarial Training, Model backbone=RoBERTa_Large2020.10 | 77.1 | 83.6 | |
| FreeLBTraining protocol=Adversarial Training, Model backbone=RoBERTa_Large2020.10 | 76.2 | 82.3 | |
| ALUMTraining protocol=Adversarial Training, Model backbone=RoBERTa_Large2020.10 | 75.9 | 81.4 | |
| VanillaTraining protocol=Standard Training, Model backbone=RoBERTa_Large2020.10 | 74.3 | 80.6 | |
| ALUM-BERT-BaseBackbone=BERT-Base, Training=Adversarial pre-training2020.04 | 63.2 | 69.8 | |
| BERT+-BaseBackbone=BERT-Base, Training=Adversarial fine-tuning2020.04 | 60.5 | 65.7 | |
| BERT-BaseBackbone=BERT-Base2020.04 | 59 | 64.8 |