We propose a large scale semantic parsing dataset focused on instruction-driven communication with an agent in Minecraft. We describe the data collection process which yields additional 35K human generated instructions with their semantic annotations. We report the performance of three baseline models and find that while a dataset of this size helps us train a usable instruction parser, it still poses interesting generalization challenges which we hope will help develop better and more robust models.