Comment by countmora

> unfortunately the tokenizer was trained on this subreddit

Do you have a source for that or was it just an assumption?