I don’t think you can “train” these LLM models. You can augment them by providing tools they can reference, but you’re not changing the underlying model.

What you’re describing - training your own model - would be a huge undertaking. You effectively want an LLM trained for next token completion, but with a narrow knowledge set.