How to choose a distribution for the likelihood step?

You can look into PyMC’s find_constrained_prior() function and see if that helps.