“State-of-the-art language models perform better when trained on a diverse dataset because there is an insufficient volume of high-quality data for any single specific task.”