Might be a silly question, so let me know if I’m going about this the wrong way.
I want to generate multiple pieces of content (30+). For instance, something close to it is generating names for a movie for a given topic.
Generating the 30+ initial movie names is fine, but then I want to rephrase them according to some rules I have.
This is where I would like to provide fine-tuning examples by giving:
- Prompt: Bad movie title before rephrasing
- Completion: Rephrased movie title that I consider good
Is it possible to create my fine tuning this way (one movie title at a time) but actually ask to rephrase 30+ titles at the same time (in a single API request)?
Will the model understand that it needs to apply the fine-tuning for each “sub-requests” in a single prompt?
Generally, the more different things you ask the model to do at one time and the more replications you request it to do, the worse results you’ll get.
It’s certainly possible it will be able to do it, but I think you’ll get better, more reliable, and consistent quality results if you discretize your requests.
With respect to the fine-tuning, the more closely your tuning examples reflect what you’re actually going to be prompting it with, the better the results will be.
So, if you want to be able to batch the re-writes through a fine-tuning model, you’ll get better results if all of your training data consists of batch requests and responses.
All that said, I suspect it will be much more cost effective to just run this through one of the existing models with a great
system message and possibly a one-shot example or few-shot examples added on.
gpt-3.5-turbo costs 8x to use what the base model costs. So, unless the fine-tuning yields dramatically better results and you plan on doing just an absurd number of these requests, you may very well do better with good prompting and an extra iteration or two.
Thanks a lot for this clear answer.
What do you mean by discretize in this context?
consistent quality results if you discretize your requests
Discretize means split them up.
(random words for minimum post length)
So I would get the best result by rephrasing the movie titles one by one instead of by batch.
That is my assumption.
It allows the model to focus only on the precise task at hand when doing each rephrasing.
When the model processes messages it looks at all the relationships between all the tokens. That’s why the computational complexity scales with the square of the context-length.
So, even though the models are very good at focusing their attention on the relevant details of the context, all those irrelevant details will exert subtle influences on the generation.
So my expectation is you will get the best quality results more consistently by processing them individually.