Creating an Instruction-to-Code model for a custom library: Strategies and Guidelines?

Hi everyone!

I need help with creating a model that can generate code for a custom library called Taipy from instructions in natural language. Taipy is an open-source Python library that generates Web Apps from Markdown f-strings with a custom syntax.

For example:
Input: “Plot the Temperature column according to the Date column.”
Output: “<|{data}|chart|mode=lines|x=Date|y=Temperature|>”

Since our library is relatively recent, I fear we need more code examples of our library for fine-tuning a model like StarCoder on our issue.

My first idea is to do few-shot learning with StarCoder and optimize the prompt using a validation set.

What do you think? Do you happen to have any strategies or guidelines that would be relevant in my case, by chance?

Thanks and regards,
Alexandre Sajus