Logit Bias not working with Fetch Requests?

I did everything to remove the word ‘assist’, yet it’s not registered correctly, somehow…

Have you tried a system message saying "Greet the user by saying “Hello, how can I help you today?”

That would work, but i specifically want to ban some certain words, starting with this…

I get this
image

Yeah, the gpt-3.5 uses cl100k_base

so the tokens are correct, it’s a big mystery to me if the API is not actually registering it OR the “assist” is made up of smaller tokens; ass-ist | as-s-ist | etc. I cant use ‘logprobs’ return either cuz they banned it in chatgpt. So really stuck here :sweat_smile:

Are you saying the OpenAI tokenizer page is wrong? :thinking:
That has some interesting implications if so…

that page is the old tokenizer, for < gpt3.5

:smile: Did not realise that! you live and learn.