Question about 5.2.3 in the paper

I was trying to reproduce the benchmark results in the paper but couldn't find any information in the paper/repo about how to perform the single-marginal inference task. It's clear for the cancer example but none of the others. Is there a deterministic leaf node that's picked for every network? 