Try rereading the whole tweet, it’s not very long. It’s specifically saying that they plan to “correct” the dataset using Grok, then retain with that dataset.

It would be way too expensive to go through it by hand