Comment by indexerror

That is an interesting observation. I wonder what percentage of the training text data for LLMs contains proper dashes, since a large part of it is user-generated content.