I tokenized 6.2mill Danbooru captions using my tokenizer; this is how long they each were, as a big long 1d tensor

initial commit