https://www.kaggle.com/competitions/google-quest-challenge/discussion/130041
RobertaTokenizer.from_pretrained('roberta-base')does not remove white space by itself. (whereas the bert-base-uncased tokenizer does- if you remove whitespace using this function, you’ll get a better score:
- def cln(x): return ” “.join(x.split()
- described here: https://www.kaggle.com/c/google-quest-challenge/discussion/129857