As far as I understood, the RoBERTa model implemented by the huggingface library, uses BPE tokenizer. Here is the link for the documentation:
