When abstracting parts of the definition of Keras Model as Models I get an underperforming network

Question

I am using an open and well know dataset "diamonds". I create a NN that predicts price (regression) and cut (classification). The architecture of the model is ra