At first glance, it appears that you are supplying unnormalized values to the p argument of the binomial. These can fall outside the range [0,1] and then since those are invalid parameters for the binomial distribution, it fires off an error. Try applying pm.math.sigmoid to squash those values into the range of valid probabilities and see if it solves your problem. The sigmoid is an example of an inverse link function which is a crucial part of every GLM. You can read more about the standard choices for link functions here.