BertWordPieceTokenizer trained on the default settings here