These are unicode exemplars.

Byte exemplars were pretokenized with the huggingface ByteLevel() pretokenizer. This is similar to the original GPT-2 approach