Hey there!
I have a file (2 pages long) that contains instructions on how to deal with the provided data and another file that contains how output should be formatted. But passing everything (instruction and output format) whenever I call API causes token wastages as well as increased output time.
Is there any way to finetune it based on the instruction document that I have so that I won’t be needing to pass same instructions with the prompt everytime?
Hi there!
The short answer is no. However, you do have an option to create a fine-tuned model which would reduce the need for instructions. Have you considered this as an option?
2 Likes
Hi @jr.2509 ,
Thanks for your quick reply!
For finetuning a model, the only approach I could fine online was to provide with multiple question-answer pair.
Is there any way to avoid that and just use document containing instructions?
Unfortunately not. You’d have to create the training data set as input for the fine-tuning process.
1 Like
What number of question answer pair should be sufficient?
It’s a bit difficult to generalize.
Personally I often start with a “quick and dirty” fine-tuned model to validate whether it works or not that contains anywhere from 30-50 data pairs. If successful, I then normally expand the data set for improved results.
I’ve had anything so far from fine-tuned models with just under 100 examples to 3,000+ examples.
If you share more details it might be easier to give further advice.
1 Like
Thanks @jr.2509 for your help!!
I’ll share details once I have some insights on this.
1 Like
What about breaking it down into smaller actions? Instead of this huge instruction set, break it down and run it through several times. You might be able to optimize parts of it with faster and with cheaper models as well. You may have the added benefit of easier upkeep and testing.