We've been chasing this rabbit since the beginning. It currently seems to be uncatchable for the use cases that would be most valuable to us - writing complex queries we've never seen before. Our use case seems to confound any notion of training or fine-tuning, since the cases we need the most help with are also the ones we have the fewest examples of.
Instead of going for generative, few-shot models, I am starting to look at the other end of the spectrum: Binary classification into deterministic query building.
With ChatGPT, you cannot realistically explain why you got some output in a way that anyone other than an AI/ML expert would find satisfying. With binary classifiers, you can precisely explain how some input resulted in some output in terms that a business person could easily digest - "You mentioned this table so it assumed you wanted to constrain on XYZ. Here's the trace from the classifiers...".
I've proposed a few schemes where you define groups of classifiers for each SQL building concern - Which tables are involved, which columns, is a join or aggregate implied, general context of business use, etc. Clearly, there are holes with this scheme, but in our domain we could plausibly fine-tune our humans to be a little bit more verbose in their use of the automagic SQL vending machine. Several hours spent training humans is probably a lot cheaper & easier than getting ChatGPT, et. al. to consistently play by our rules.