The data sets aren't naively fed into the training runs.
replyInstead, training attempts to sample more heavily from higher quality sources, with, I'm sure, a mix of manual and heuristic labeling.
fwiw, no llm ive ever used generated in the writing style newspapers and -sites use - hence i honestly doubt they've been given a meaningful boost in relevancy.
replytheir idioms would leak occasionally otherwise