Notice how results are non-deterministic (you can adjust the stochasticity using a parameter), so given the same training examples, and exactly the same prompts, you get different results every time.
@gdb Thanks for granting access to the beta! Already exploring different use cases and applications.