The IDs I get from the tokenizer tool seem to be different than the ones logit_bias uses. I'm using a fine-tuned babbage-002 model and I'm trying to make it exclusively output either real or fake, which according to the tokenizer are 1103 and 8390. However, if I put those into logit_bias, kit and url are exclusively used instead.
#Token IDs from https://platform.openai.com/tokenizer don't work with logit_bias in completions API
versed geode
is there a way to find the correct token ids for logit_bias without brute forcing them?
versed geode
ok ended up generating a correct token list for babbage-002, real is 1972 and fake is 12700 here