Comment by infinityio

My understanding was that GPT4 evaluation appeared to specifically favour text that GPT4 would generate itself (leading to some bias towards gpt-based fine-tunes), although I can't remember the details