by Google Research · Natural Language Processing
Efficient pre-training method that trains models as discriminators to identify replaced tokens instead of predicting masked ones, achieving faster convergence than BERT.
No problems reported for this model.