Classification on Crowdflower economic news relevance (test)
82.5AccuracyBERT_BASE Adapters
Evaluation Results
| Method | Links | |
|---|---|---|
| BERT_BASE AdaptersBackbone=BERT_BASE, Evaluation Protocol=Adapter-tuning, Total number of parameters=1.19x, Trained parameters per task=1.14%2019.02 | 82.5 | |
| BERT_BASE Fine-tuneBackbone=BERT_BASE, Evaluation Protocol=Fine-tuning, Total number of parameters=17x, Trained parameters per task=100%2019.02 | 82.1 | |
| No BERT baseline2019.02 | 81.1 | |
| BERT_BASE Variable FTBackbone=BERT_BASE, Evaluation Protocol=Variable Fine-tuning, Total number of parameters=9.9x, Trained parameters per task=52.9%2019.02 | 78.9 |