Training a neural net with Rhino command history

If I get this right, all current efforts for ML assisted 3D model generation are polygon/point-cloud based and work similarly to stable diffusion.
To build nurbs based and possibly parametric geometry, would it be a viable approach to use rhino command history or even the full code history as training data in a language model with either text based classification or corresponding images as classification?
Do these approaches exist already? Where to start, and what could be the problems?