The part I am most interested in is how much of it is on device and how much of your data is trained only on your own model vs hovered up into a larger model.

On device, my own model. Great.