Wi-Fi CSI Presence Detection: Dataset and Pre-trained Models

This dataset contains WiFi Channel State Information (CSI) records collected for device-free binary human presence detection in a residential environment using two ESP32-S3 microcontrollers. Captured over a 2.4 GHz ESP-NOW dedicated link with 40 MHz bandwidth (HT40) at a 30 Hz nominal sampling rate, the dataset is structured into raw and processed directories. Version 2.0.0 Updates:The dataset has been significantly expanded from the initial pilot to include the main experimental campaign (multi-position static occupancy) and a cross-domain generalization campaign testing robustness against varying distances and environments. The raw folder now houses continuous CSV time-series records and JSON metadata for all three campaigns. The processed directory contains the finalized tabular dataset (CSV and Parquet formats) yielding non-overlapping 2-second windows, alongside pre-split train, val, and test sets to ensure standardized benchmarking. Each window initially comprises 648 temporal statistical features extracted from 162 valid subcarriers. Additionally, this release includes a comprehensive model registry featuring pre-trained pipelines (Random Forest, Gradient Boosting, SVM, and MLP). Models are provided in both Full (648 features) and Lightweight (20 features across 5 optimal subcarriers) variants, achieving up to 96.9% feature reduction with minimal macro F1 degradation. All models include serialized pipelines, scaling artifacts, and standard Model Cards. This work was designed as part of an undergraduate engineering thesis at UFMG (Federal University of Minas Gerais, Brazil).