Generating unwanted answers in Fine-tuning

API
,
1

Good evening

I am facing a challenge with the GPT-4O-MINI model that I have customized using Fine-tuning technique on a specific dataset in (.jsonl) format. Even though I trained the model on specific data, it is not producing the correct expected queries, instead it is generating unwanted answers or queries.

Are there any specific settings or parameters during the customization or generation process that the model relies mainly on the data provided during customization, so that it gives priority to this data and does not rely on its general knowledge except only when the user asks a question from outside the trained data, answers are generated from outside it?

2

This symptom has several reports recently when using gpt-4o-mini, to where you’d think fine-tune was non functional and the attempts degraded the model in general.

Try the fine-tune file against gpt-3.5-turbo or gpt-4o, and see if it doesn’t magically work better, to gather evidence that gpt-4o-mini is being worthless for fine-tune.

You can also check the report to get training loss. Download and decode from base64.

3

Can you add an example? Of unwanted and desired.