Predicting a computer users covert cognitive state, such as attention, has previously proven to be difficult, as cognitive states are induced trough complex interaction of hidden brain processes that are difficult to capture in a traditional rule-based methods. An alternative approach to modeling cognitive states is through machine learning, which however, requires that a wide range of data is collected from the user. In this paper, we describe our software for collecting a wide range of data from office workers during everyday computer work. The data collection process is relatively unobtrusive, as it can be run as a background process on the users computer and does not require extensive computational resources. We conclude by discussing practical issues, such as data sample frequency, where one wants to strike a balance between good enough data quality for machine learning and unobtrusiveness for the user.