Multiple strange spelling errors with

Prompting
,
1

Screenshot 2024-07-15 at 9.41.58 AM
Screenshot 2024-07-15 at 9.41.58 AM574×1037 42.3 KB

(I use google firestore to store a log of my API calls inputs/outputs)

Model used: gpt-4-turbo
Settings: all default

Very strange output issue, as you can see the spelling of “Book Promise” is incorrect multiple times in a row…

Relevant part of my prompt:

// lots of other instructions about how to output the content, then I add formatting instructions as follows

CRITICAL FORMATTING INSTRUCTIONS

Ensure to *strictly* follow the following structure in the generated output:

BOOK PROMISE: [the book promise goes here].

READER PROFILE: [the reader profile goes here].

... etc.

For the life of me I can’t figure out what keeps causing this very specific output spelling issue… I have the strict output so that I can use regex to match the relevant text I need from the output, so spelling does matter (as is usually the case anyways).

Thoughts?

2

Some things:

  1. We would need to see the full instructions. Not just a snippet.
  2. Are you using logit_bias, frequency_penalty, or presence_penalty
    ** Even better, post your full parameters here
  3. Why are you using regex? Why not just use json_mode and then grab the item by key?