GPT-3.5 probably has substantially lower cross-entropy loss before instruction fine-tuning as well.

Meta’s LLaMA v. GPT-3 comparisons are to OpenAI’s 2020 release of GPT-3, but there’s been significant progress since then.