Hey all
Novice to the world of prompting and AI here! I have recently started testing OpenAI’s models with a site builder project.
My end goal is to have AI automatically generate a website for users based on a prompt. I have chunked this goal into three portions, based on the logic of our builder:
- A website consists of pages.
- Each page consists of a header, footer, and any number of sections.
- Within sections, users can insert blocks such as a text, button, accordion and so on.
Based on the above logic, I started with testing whether the model can auto-generate sections. However, despite my best attempts at crafting a well-formatted prompt, the returned sections are not quite there yet. It’s also a hit and miss, as (more often than not) the returned section causes an error.
I am considering fine-tuning a model to achieve better results.
Fortunately enough, we created roughly 50 premade sections (.tsx files) which could be use as training data AIUI. Optionally, we can further enhance this by using a dozen or so page templates, as well as a dozen site templates.
My two questions are as follows:
- Which model is best suited to be fine-tuned in my scenario? The returned sections should be formatted as a Typescript Object Literal, so should I be using a “Structured outputs” model?
- Am I correct in saying that I need to create a JSONL file consisting of a prompt and returned value? If so, are 50 sections enough for this, or would you advise more/less?
Finally, if you have any other advice here, I’d love to hear your feedback!
Thank you in advance!