Some questions about the RoBERTa-wwm-ext-large

[hotchilipowder]

Hello, It's a nice work for developers.

I am using these pre-train-models to some tasks. It works in the BERT-wwm-ext with Huggingface's pytorch-bert.

As you said,

文本不需要经过分词，wwm只影响预训练过程，不影响下游任务的输入

However, when I use the RoBERTa-wwm-ext-large. I have a question.

In Huggingfac's documents,it looks like we need two files: vocab_file, merges_file

How to deal with it? Just use the Bert's tokenizer? Thank you

[ymcui]

Load our RoBERTa(-like BERT) using BERT's interface.