This idea of generating data from seeds sounds wrong: it creates lower quality data than human conversations. I think a better approach would be to retro-engineer humanly written text to get instructions. The main question is are instructions more important than outputs? #llama

→ View original post on X — @maximelabonne