Describir: Evaluating Foundational Models Using Insights From Their Pretraining Data