I disagree. I'd include overfitting for LLMs as creating unreasonably strong connections to individual sequences used for training, whereas a good mix of that a

I disagree. I'd include overfitting for LLMs as creating unreasonably strong connections to individual sequences used for training, whereas a good mix of that and connections between chunks of those sequences are required.