[Veedrac]

There are many BERT-based models that would have made for a good numeric comparison, had they tested on few-shot learning, but I'm not aware of any that have.

[msamwald]

Well, in table 1 they compare to RoBERTa trained in a vanilla supervised fashion?