Context Behavioral Context refers to a wide range of attributes describing what is going on with you: where you are (home, school, work, at the beach, at a restaurant), what you are doing (sleeping, eating, in a meeting, computer work, exercising, shower), who you are with (family, friends, co-workers), your body posture state (sitting, standing, walking, running), and so on. The ability to automatically (effortlessly, frequently, objectively) recognize behavioral context can serve many domains. Medical applications can monitor physical activity or eating habits; aging-at-home programs can log older adults' physical, social, and mental behavior; personal assistant systems can better server the user if they are aware of the context. In-the-wild (in real life), natural behavior is complex, composed of different aspects, and has high variability. You can run outside at the beach, with friends with your phone in the pocket; you can also run indoors, at the gym, on a treadmill, with your phone motionless next to you. This high variability makes context-recognition a hard task to perform **in-the-wild**. Content The ExtraSensory Dataset was collected from 60 participants where each person participated approximately 7 days. We installed our data-collection mobile app on their *personal phone* and it was used to collect both sensor-measurements and context-labels. The sensor-measurements were recorded automatically for a window of 20-seconds every minute. This included accelerometer, gyroscope, magnetometer, audio, location, and phone-state from the person's phone, as well as accelerometer and compass from an additional smartwatch that we provided. In addition, the app's interface had many mechanisms for self-reporting the relevant context-labels, including reporting past context, near future, responding to notifications, and more. The flexible interface allowed to collect many labels with minimal effort and interaction-time, to avoid interfering with the natural behavior. The data was collected in-the-wild: participants used their phone in any way that was convenient to them, they engaged in their regular behavior and reported an combinations of labels that fit their context. For every participant (or "user"), the dataset has a CSV file with pre-computed features that we extracted from the sensors and with labels. Each row has a separate example (representing 1 minute) and is indexed by the timestamp (seconds since the epoch). There are columns for the sensor-features, with the prefix of the column name indicating the sensor it came from (e.g. prefix "raw_acc:" indicating a feature came from the raw phone accelerometer measurements). There are columns for 51 diverse context-labels and the value for an example-label pair is either 1 (the label is relevant for the example), 0 (the label is not relevant), or 'NaN' (missing information). Here, we provide data for 2 of the 60 participants. You can use this partial data to get familiar with the data and practice algorithms. The full dataset is publicly available at http://extrasensory.ucsd.edu. The website has additional parts of the data (such as a wider range of the original reported labels, location coordinates, mood labels from part of the participants). If you use the data for your publications, you are required to cite our original paper Vaizman, Y., Ellis, K., and Lanckriet, G. "Recognizing Detailed Human Context In-the-Wild from Smartphones and Smartwatches". IEEE Pervasive Computing, vol. 16, no. 4, October-December 2017, pp. 62-74. Read the information at http://extrasensory.ucsd.edu and the original paper for more details. Acknowledgements The dataset was collected by Yonatan Vaizman and Katherine Ellis, under the supervision of prof. Gert Lanckriet, all from the department of Electrical and Computer Engineering, University of California, San Diego. Inspiration The ExtraSensory Dataset can serve as a benchmark to compare methods for context-recognition (or context-awareness, activity recognition, daily activity detection). You can focus on specific sensors or on specific context-labels. You can suggest new models and classifiers, train them on the data and evaluate their performance on the data.