Contributed to RoBERTa, showing that longer training on more data with tuned hyperparameters substantially improves BERT.
Works in the archive (1)
“Lead” marks the first-listed author — a byline fact, not a measure of influence; co-authorship is shown as recorded.