Usually for these adhoc contests, once my heuristics reach a state that I can't improve my solution anymore, I explore the tests and finetune constants like max iterations or leaves to get the highest score, and rerun many times to average the points. For this adhoc contest, not all tests are released so I submitted multiple times and estimate what the tests on the server covers.

I don't think my approach for closed-set test cases adhoc contest is the proper approach, it's hacky but yeah, that's what I do for this one. For open-set test cases, you only submit the solution so it's a bit easier