We all know these models are trained on the internet. And the internet is full of spam, duplicates, toxic text, and personal data nobody should be memorizing. So how does any of that turn into a model that actually works? It doesn't go in raw. The data runs through a pipeline