Comment by winddude

Seems to really fall apart on subsequent prompts, and a few times I've had code end up in the "thinking" tokens.

I'm guessing most of the training data was single-turn, instead of multi-turn, but that should be relatively easy to iterate on.