[andai]

I forget which model (LLaMA 3?) but I heard 95% of the training data was English.