Visual Madlibs contains 360,001 natural language descriptions for 10,738 images. The dataset uses automatically generated fill-in-the-blank templates to collect descriptions of several targets, including: people and objects, appearance, activities and interactions, and inferences about general scenes or broader contexts.