Greedyness parameter

Hej guys, thank you for your amazing work!
I just have one clarification question: is there some sort of heuristics to decide which value of greedyness one should use when running the model? It is 0.5 by default, but it does not seem to capture some important co-references when I run it on my data. So I assume that this value has to be adjusted empirically for every unique data that we run neural coref with.
However, it would be nice if there was some sort of a guide that tells which values of greedyness output which results. Is there anything like that? Or someone can share their personal experience with setting this parameter?

Thanks.