Evaluating GPT-3 and GPT-4 on the Winograd Schema Challenge

I'm guessing they trained it on most if not all of those datasets.