Comment by conradkay

My guess is that it doesn't remove that much of the data, and the post-training data (not just randomly scraped from the web) probably matters more