Must match the tokenizer vocab. size, i.e., 50265.

Others also mentioned this: https://discuss.pytorch.org/t/hugging-faces-roberta-config-and-tokenizer-do-not-have-matching-vocabulary/134868