I’m working in a related area and I’m rather curious about this point. In what way is GPT-4 overfit? Does overfit in this context mean the conventional: validation loss went up with additional training, or something special?