Microsoft found that small language models can exceed the performance of much larger ones when trained to specialize in a single area. Researchers fine-tuned the Mistral 7B model to create Orca-Math, ...