I am retraining Roberta and want to evaluate the MLM accuracy to check the difference before retraining and after retraining. I know that we can use a data collator during retraining for generating masked input.
But for evaulating, if I have a selected set of 100 sentences, how can I generate a masked version of these sentences?
For example, if the sentence was: This movie was very interesting to watch
I want to have a sentence that looks like this: This movie was very ###### to watch.
How to do this?
Thanks in advance!