I've been experimenting with Magpie-style instruction data extraction using Llama 3.1 70B, and noticed kind of an odd result: 13% of the responses generated were as if the user asked a question…

Posters: xlr8harder. xlr8harder

4 likes recorded by the source corpus.