Why do some tokenizers assign negative floats to each token? For instance, I am looking at this json file, and the tokens start about 1/3 of the way down the page. Each one is part of a two-element list with the structure "[<token>, negative decimal number with 15 digits of accuracy]"