Looks like llama 7b was trained on 4 times more tokens.