i mean, i’m not arguing for this position necessarily. i agree with you.

that said, in practice, the “bad data” comes out in the wash. for example, all of Reddit is likely in their datasets. who knows what their data cleaning process is like, but diversity in the data is generally a good thing from a training perspective.