Has someone been able to fine-tune the GPT 3.5 model to recognize a JSON schema by an assigned name, and then later, use just that schema name in prompts to obtain structured output rather than including the entire JSON schema in prompt?
Are you using the response_format={‘type’:“json_object”} in the latest gpt-3.5-turbo-1106 model? I have found it to be pretty good at understanding what I want with a light description of the key/values of the JSON schema. Obviously not exactly what you asked for, since I do mention something about the schema in each prompt and may not work for larger/complex JSON schemas. But would be super interested in seeing if a fine-tuned version would just remember the assigned name. You could use the json_object response_format to actually generate the fine-tuning data if you needed more training examples.
1 Like