Why Meta trained a small 8B model on 15 trillion tokens, and why that's cheaper than it sounds