Comment by SwellJoe

I think they're distilling "reasoning", not merely code. There's plenty of human generated code. What they're trying to extract is the process by which really large models "think" their way through complicated problems. That's what all the "traces" datasets on HuggingFace are about.