databrickslabs/dolly ? reverse-engineered prompt

Build me a simple repo for Databricks Dolly that lets me fine tune and run the model from a notebook or script.

I want it to include a training flow for the Dolly instruction tuned language model, using the provided instruction response dataset, and save the trained model when it finishes. It should be easy to switch between different GPU setups like A100, A10, and V100, and it should work with the kind of setup Databricks recommends in the README. Please also include a small example for generating responses from the model after training, so I can ask it normal questions and get answers back.

Make sure the code is clean and runnable, with sensible defaults and clear comments where someone would expect them. If you need to check the latest docs for any library behavior, go ahead and look them up online.