Tl;DR: It's quite a lot worse than LLama 13B in every way that is measured. It benchmarks between GPT-J (6B) and GPT-NeoX-20B, although is worse than GPT-J (6B)

Tl;DR: It's quite a lot worse than LLama 13B in every way that is measured. It benchmarks between GPT-J (6B) and GPT-NeoX-20B, although is worse than GPT-J (6B) in a few benchmarks. There are no good benchmarks for text generation, but it would be surprising if it was competitive with the best open models. It's unclear how much RLHF can help here - AFAIK no one has tried out DataBrick's Dolly to compare a GPT-J based RLHF model.