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