Modern smart grids operate as cyber-physical systems in which tightly integrated communication, sensing, and control infrastructure increases both operational capability and cybersecurity risk. Machine-learning-based attack detection has emerged as a promising defense mechanism; however, research in this area is constrained by limited availability of labeled cyber-physical datasets and the high cost and complexity of existing real-time testbed infrastructures. This paper presents a low-cost, reproducible cyber physical smart grid testbed designed specifically for structured dataset generation under both normal and attack-induced operating conditions. The testbed integrates distributed voltage sensing, relay-based switching, MQTT-based communication, and scripted cyber attack simulation to emu-late measurement-and-control interactions within a simplified distribution feeder. Controlled experiments were conducted to generate labeled telemetry representing normal operation, false data injection, denial-of-service, and power theft scenarios. The resulting dataset was used to evaluate three supervised machine learning models: Random Forest, Support Vector Machine, and XGBoost. Experimental results demonstrate that the proposed testbed can generate sufficiently representative cyber physical telemetry to enable effective attack classification. By emphasizing affordability and reproducibility, this work lowers the entry barrier for dataset-driven smart grid cybersecurity experimentation.