Code generation and fine tuning solutions

So as we all know the only way today to work with code generation (text to sql) is now Turbo GPT, since CODEX is gone, which is slow and make lots of code syntax errors.
The way we worked with it , is describing the scheme each prompt (e.g table(col1, col2) ), then insert the relevant user prompt after the scheme every time.
Because we had to insert the table structure every time, and because the model is ridiculously slow, we tried to fine tune the Davinci.

The way we fine tuned it is by inserting text prompts and their corresponding SQL code, with a huge high quality list, and trained the model.

Then we tried to run the trained model and expected faster better results.

What we discovered is that the model is almost as slow as before, and now it even makes more weird syntax errors, and the whole expensive training was useless.

We tried to train the Curie as well with no success.

Currently the only option is to insert the full table into each prompt, and wait 10-30 seconds (because they intentionally limit the api), and get responds that are 30% of the times wrong.

Are there any other ways to train these models by just inserting them the table structure so that we do not have to input this every prompt? or, at least train them and force them to actually listen to our trainings ?

We went above and beyond with this, and we could not make this work, the generated code is slow and wrong.

Suggestions ? Thanks.