Can finetuning a model with excessive dataset can make the model smarter

I’ve got a scenario where the model needs to generate JSON for a given input. The JSON should be in a specific format with particular keys. I’ve prepared a dataset with 50+ examples. However, the model is not generating very good results. Moreover, I’ve not covered all the cases yet.
I was wondering if I created too many examples. Would that make the gpt-3.5-turbo model any more intelligent and smart with better memory?

Hi and welcome to the Developer Forum!

Show the model a single example json output in the format you want with placeholder entries in the fields then put it in ### makrers top and bottom and then tell the model to use that as the output template, should fix your issue.

sorry, I’m not getting you exactly.

For your reference, the following is an example from my dataset

{"messages": [{"role": "system", "content": "Given the description for a video or image, generate a JSON."}, {"role": "user", "content": "A video that is 1920x1080 and 3 seconds long, with the text 'This text is only visible for one second' in white on a black background. The text should only be visible for one second."}, {"role": "assistant", "content": "{\"output_format\": \"mp4\", \"width\": 1920, \"height\": 1080, \"duration\": \"3 s\", \"elements\": [{\"type\": \"text\", \"track\": 1, \"time\": \"0 s\", \"duration\": \"1 s\", \"x\": \"50%\", \"y\": \"50%\", \"background_color\": \"#000000\", \"fill_color\": \"#ffffff\", \"text\": \"This text is only visible for one second\", \"font_family\": \"Open Sans\"}]}"}]}

In the next examples, the attributes keep adding up.

Doubling the no. of epochs significantly improved the results for me.