The Image Paragraph Captioning dataset can be used to evaluate description snippets generated for images. The dataset contains 19,561 images from the Visual Genome dataset. Each image contains a paragraph. The training/evaluation/test sets contain 14,575, 2,487, and 2,489 images, respectively.

Each image also contains 50 region descriptions (phrases describing a specific part of the image), 35 objects, 26 attributes, and 21 relations, as well as 17 question-answer pairs.