Epidemiology of Large Language Models: A Benchmark for Observational Distribution Knowledge