When I fine-tune the latest gpt-4o and use the response_format={“type”: “json_object”}, the fine-tuned model outputs lots of total nonsense with lots of random keys and values. It was reported before, but for some reason this https_community.openai.com/t/inconsistent-fine-tuning-results-gpt-4o-vs-gpt-4o-mini/947573/10 is still marked as “resolved”, even though it has not been resolved (sorry for not pasting the full link properly: it says I cannot do it. So remove the ‘https_’ it will work). In fact, the issue remains to be there for quite a while now, and more people have reported the same thing now.
To make it all simpler to investigate, I have put together a little Jupyter Notebook (Google Colab) demonstrating the issue:
- Jupyter Notebook: https_colab.research.google.com/drive/1-juShZWnbna1i3HkJW1y_HPqcmfoEvWn#scrollTo=Lm0wVS3HiYsy
- Data: https_docs.google.com/spreadsheets/d/1A532tiaMggBMvC-4mTU1t7x51GBBzKXF2zLtulFOWDk/edit?gid=0#gid=0 (just a bunch of publicly available problems and solutions, and some are written by me)
In order to run the notebook, please make a copy of it, add your own api_key and upload the data. It is not much data, but it is enough to see the issue (and when I run it with much more data, everything gets only worse).
It would be great if this gets looked at. I am sorry if I missed another forum where this got answered, please direct me there.