I read that they training data was not scrapped from the internet but generated through meta llama model. Maybe this is the reason behind such behavior?