Why were you able to change the eos and bos token IDs from 1 to 11 in the config.json without replacing the tokenizer? Shouldn't it be consistent with training?
11 is the correct id, it was wrong before
· Sign up or log in to comment