Hi,
I read the getting started documentation and understood how the fine tuning works with the format:
{"prompt": "<prompt text>", "completion": "<ideal generated text>"}
Q1: I assume codex does not support the type of code I want to generate. In this case do I resort to fine tuning to teach a new programming language (PL)?
Q2: Assuming fine tuning is the way to go: What is the best way to create a model in terms of complexity? E.g. let’s say I am teaching HTML. Do I go with complexity level 0 or 10 or somewhere inbetween, where complexity 0 means I enter:
{"prompt": "give me a link that points at google that says hello",
"completion": "<a href="https://google.com">Hello</a>}
Do I do little snippets like this? Or go for Complexity 10, e.g.
{"prompt": "give me an html page for an ecommerce website
optimized for selling XYZ", "completion": "<entire html + css + javascript>"}
Or do I go inbetween?
Q3: Can you teach to mix languages? E.g. instead of a link, I want a dropdown html element that includes both javascript and css classes to describe. How can I do that?
SaaSBox is an application that hosts front-ends for SaaS companies. The frontend pages have a well defined problem space (3-5 pages with specific content/functionality). The quirk is there is a mix of languages and tags to be generated. We code these by hand, but I can easily see it all being generated by AI. It would be great if anyone is interested to work on this as part of a research, or explain how big of a problem this is.
Cheers.
Team SaaSBox