Descriptor: The WiFi-Based Human Activity Recognition Under Everyday Life Changes Dataset (WiFi-Wild)

WiFi-based Human Activity Recognition (HAR) provides a non-invasive alternative to camera- and wearable-based sensing, but model performance often degrades under environmental and subject-specific variations. This work introduces the WiFi-Wild dataset, a publicly available collection of WiFi Channel State Information (CSI) signals designed to capture everyday life changes such as furniture rearrangement, transmitter/receiver placement, device interference, and subject diversity. The WiFi-Wild dataset was collected in a residential environment using ESP32 devices, with a single participant performing a range of daily activities under multiple spatial and environmental scenarios, and additional participant appearances were included to evaluate subject variability. To demonstrate dataset utility, a hybrid deep learning model combining a Convolutional Neural Network (CNN) with an attention-based Bidirectional Long Short-Term Memory (BiLSTM) network was trained for activity classification and evaluated across different configurations. Results show that environmental and subject-related variations substantially affect model performance, with furniture rearrangements and TX/RX repositioning reducing F1-scores by up to 21%, and signal interference or clothing changes causing decreases of 7–12%. The WiFi-Wild dataset enables systematic, reproducible evaluation of WiFi-based HAR models under realistic conditions and supports the development of more robust systems.