I have a relatively large dataset of example prompts,and their corresponding assistant outputs which I was going to use to finetune a model which uses the chat completions API.
The main reason for this, is that I was using function calling, and the model would often produce outputs that didn’t match the JSON schema.
I have since decided to switch to the Assistant API due to the added context features and window.
Is it still worth finetuning the model?
Personally, my workflow is to test the model “as-is” then if there are issues, I use prompt design to attempt to solve them, if I need domain specific information from large datasets, I’ll use RAG. If at that point I still have output formatting problems, I will first try programmatic correction with traditional code and as a finial step I may choose to fine-tune a model.
This works for my use-cases but may not be suitable for all, however, I would still say that fine-tuning is primarily a finial step when polishing a models output if required, of course if you are doing something like moderation, or specific classification, then you may want to go for fine-tuning right away.
Would you say that given I already have a dataset ready I “might as well” finetune it? Or should I leave it for now.
By all means, give your fine-tune a try, AI and it’s related subsystems are too new a technology for anyone to be able to say categorically that Dataset A will produce outputs of a given type B with much certainly, we can give general findings and rules of thumb, but anything over and above that will, a) probably be wrong and b) be out of date in 6 months.