I have fine tuned the 4o mini model, with the minimum number of examples. It is supposed to generate json files in different structures.
All of the examples have about 50-70 elements in them, however, when i call the api, or play with it in playground, it just runs forever, it never occurs to stop. around 400 or so elements it start over with a new json, but it just doesnt stop.
I have even tried adding this: \n\n###\n\n to the end of all my fine tune example messages, and specified this as the stop sequence for the assistants system prompt.
I did not give any validation examples to the fine tuning, would that solve the case, or is it some other thing?
Any ideas why this might happen? Thanks in advance!
You’ve likely over-fitted the AI on producing JSONs in a pattern. You powered-out the intelligence behind the generation.
After creating long enough of a list, combined with examples of very long repetitive JSON with only one “chat stop” per training example, you get an AI where the next thing to output is always more list: "value",, never "value"} nor end-of-message.
What I would do is make an “index” key to go along with each object, so you can instruct and the AI can track for “produce an array of 10 objects”.
Otherwise, you can take a look at what natural stop sequences arise after a JSON, so it doesn’t keep starting more. Your own stop sequence never seen shows that ending the output isn’t even on the AI’s mind.
I would suggest the frequency_penalty parameter, but these along with logit_bias seem broken and non-functional.