The data sets aren't naively fed into the training runs.

Instead, training attempts to sample more heavily from higher quality sources, with, I'm sure, a mix of manual and heuristic labeling.