Opening up of spectrum for shared use, such as the Citizen Radio Broadband Service (CBRS) band, offers unprecedented opportunities for allowing commercial operators to operate in frequencies otherwise reserved for federal use only. Specifically in the CBRS band, the challenge of detecting the highest priority incumbent radar reliably forces severe restrictions on the transmit power for operators deploying LTE networks. While Machine Learning (ML)-based solutions have demonstrated the potential for detecting weak radar signals in fully overlapping secondary signals, there exists a fundamental gap in porting these methods for practical, real-world conditions due to a key reason: There are no accessible data-sets or even controlled methods to generate such datasets today over-the-air (OTA), where radar and LTE ‘overlap’ in a number of challenging SINR conditions. This article makes three contributions: It describes the first publicly available CBRS over-lapping and non-overlapping LTE and radar OTA dataset in the 3.5 GHz band using an experimental testbed composed of software defined radios; It describes the first-of-its-kind open source Application Programming Interface (API) that can configure automatically multiple transmitters and receiver radios, synchronize them, remove the Tx local oscillators-induced artifacts, and carefully set their parameters such as sampling rates, center frequencies, and time duration for sample collection, ultimately resulting in high-fidelity data in the Signal Metadata Format (SigMF); It demonstrates the utility of the CBRS dataset by adapting the well-known ML model called “You Only Look Once” (YOLO) for detecting and localizing the radar and LTE signals with near-perfect accuracy, pointing to the possibility that current FCC-mandated power thresholds can be lowered for cellular operators in the CBRS band.