“Meta's Llama 3 model was trained on 15 trillion tokens from a total available Common Crawl dataset of 30 trillion tokens.”