Hi @ricardoV94 - are we talking about the code I posted above, in the beginning of this thread?
If so, that code is the model, the priors, and the toy observed-data. At least with my setup, copying and pasting that code recreates the errors. If you re-run the whole code block multiple times (including toy data generation), it should behave as described, sometimes “working” and sometimes throwing a bad initial energy error.
The only thing that changes between “successful” (= no error returned, even though the chains don’t converge) and unsuccessful (error message = “bad initial evergy, ypred -inf”) is the observed data yy. I think this is because yy has some random error on it that changes a little every time, if you rerun the whole block of code including the toy data generation. The bad initial energy errors are usually thrown early in the sampling or not at all.
But if you’re talking about the exponentiated model, I’m still tinkering with that…