Which kind of BERT version is used?

[lanzhuzhu]

Thanks for your open sourcing. This work is really solid. I have some questions about the implementation.

Which kind of BERT is used in your paper? As we know BERT_LARGE or BERT_BASE could bring a quite different performance due to the number of parameters.

Even though the default setting is BERT_BASE, it would be better to clarify the used version of BERT somewhere in your paper or Github. I asked this because I fine-tuned the BERT_BASE_UNCASED model as a relation classifier but the final F1 score on the original test dataset is 0.6128, much lower than SpanBERT or KnowBERT. The result of BERT RE model is not clear in your work, so it is hard to compare.