Getting fined tuned 3.5 to NOT speak and just give specific info

I’m trying to make it give me a table of information based on what I give it(for my specific app). Has anyone had a any luck tuning the API to not speak and only do it’s given task?

It works for the most part but I see the occasional “Here is your table” before getting the table. Mind you I only trained it on ten JSONL objects. Perhaps I should use more?

you mean like this ?

Caould you give a sample of what you used as a training example ?

The prompt you used and the training sample, if structured correctly irrespective of the number should be able to effect the format in which your output is generated

Try ten thousand… :laughing:

  • Have a specific unique system prompt used in both fine-tune and in using your model,
  • Have ample coverage of examples for all types of inputs that must have this behavior, making inference for the middle cases easy.
  • have a user instruction code word repeated for your “no output just table” production.
  • even if you train it on making tables just as you like, expect regular “turbo” training when you venture away from the trained task.

finally, see if you can get the output with a whole bunch of prompting that is 1/8th the cost of using a fine-tune model.