Japanese RoBERTa base size trained by wikipedia dump 20220905 using fairseq.
Tokenizer is japanese_roberta_tokenizer.