We introduce the Something-something V2 dataset, which contains captions of finely-varying human-object interactions. We also discuss various baseline models, and show that neural networks show surprisingly strong performance on many of the very hard, detailed discrimination tasks associated with this dataset.