Hello Bayes,
It’s not entirely clear to me what you are attempting to do. It looks a bit like logistic regression, but not exactly. You can check this example to see if it’s of any use.
To your main question, you are attempting to estimate 20 parameters (10 betas and 10 alphas) from 10 data points. Except in specialized circumstances, this is unlikely to work out well and, I assume, is not what you intend. If I remove the , shape=shape from the declaration of both beta and alpha (and alter swap out the truncated normal with a beta), I sampling is faster and has fewer (but not zero) divergences. To comment further, I would need to know more about the intent of the model.