I'm very surprised to see this considering the models aren't chat-tuned and shouldn't be used for chat, at least, by the disclaimer Meta provided. I wish they

I'm very surprised to see this considering the models aren't chat-tuned and shouldn't be used for chat, at least, by the disclaimer Meta provided. I wish they didn't include the disclaimer just for the sake of having one, they provided a chat format, so they must have trained _some_ chat, and we can see here its good enough at chat to not act like a pure LLM, or even close. Labelling it with a disclaimer it doesn't need causes general confusion in scenarios like Phi-2, which had no chat tuning but great scores, so people couldn't understand "why it didn't work".