“The Llama 3 model was reportedly trained on 15 trillion tokens.”