Dataset: Less is More: The Dilution Effect in Multi-Link Wireless Sensing

We release a comprehensive WiFi Channel State Information (CSI) dataset collected from a 9-node ESP32-C3 mesh network deployed in a 5×4 meter residential environment. The dataset comprises over 300 hours of continuous recordings across 72 directional RF links, totaling 1.6 million time windows with semi-automatic ground truth labels for presence, activity, and occupancy count. Each sample includes 52 subcarriers of complex-valued CSI measurements at approximately 10 Hz per link. The dataset supports research in presence detection, activity recognition, device-free localization, and sensor fusion, and will be publicly available on Zenodo under an open license. Version 2 (2026-08-20): adds the ground-truth labels and their provenance documentation. See README_LABELS.md first: the primary labels are an hourly retrospective occupancy diary (312 h); the ESP32-CAM ground truth is per-frame JPEG-size metadata over ~33 h, captured so that no imagery was ever stored (privacy by design in a lived-in apartment - no images or video exist in this dataset). The README also documents a known hour-of-day structure in the labels that analyses should control for.