Activity recognition, e.g., identifying individuals, recognizing their physical activities, or estimating their number in a room, based on WiFi’s Channel State Information (CSI) has been studied intensively in the last decade. While most existing works consider analyzing CSI data from a single person in a rather constrained environment, almost none of them has been successful in generalizing these results to unconstrained, real-world environments, in particular, when multiple individuals are present. In this paper, to address this problem, we introduce a fully annotated dataset ($\approx$ 70 GB of data) containing CSI and environmental data collected from two real-world offices over multiple days of continuous monitoring. To the best of our knowledge, this is the first freely available dataset of its kind. On the one hand, our dataset evidences that vastly disregarded {\em implicit changes} in the environment – due to small objects being repositioned, added or removed – are the main reason for the lack of generalizability by existing approaches. On the other hand, we expect it to promote further research work in this area and, thereby, to facilitate general solutions for CSI-based activity recognition in real-world environments.