Syn2Real is a dataset in the synthetic-to-real vision domain. The dataset is designed to encourage further development of strong domain transfer methods. The goal is to train a model on a synthetic "source" domain and then update it to improve its performance on a real "target" domain without using any object annotations. The dataset includes three tasks: a more traditional closed-set classification task with a known set of categories; a less studied open-set classification task where the object categories are unknown in the target domain; and an object detection task that involves localizing object instances by predicting their bounding boxes and corresponding class labels.