Levels of state repression and the frequency, severity, and targets of human rights abuses vary spatially within states. However, most previous studies on these topics have only considered repression in the aggregate. This is problematic because it ignores variation in institutional structures and decision-making processes within countries. We introduce a novel approach for measuring subnational levels of repression using machine learning and human coding techniques. We apply hand coding, dictionary-based approaches and supervised machine learning methods to extract and code physical integrity rights allegations from annual country human rights reports. We present preliminary descriptive statistics from a working georeferenced pilot version of the dataset for the period 1999-2016. We plan to continue to update this dataset.