Building AI Testbed Suite (BATS): Physics-Grounded Sensor Time-Series with Knowledge Graph Annotations and Auto-Gradable QA Benchmarks

This dataset contains 18 complete Building AI testbeds generated from three U.S. DOE commercial prototype building models (small, medium, and large office, ASHRAE 90.1-2019 edition) across three ASHRAE climate zones (2A Tampa FL, 4A New York NY, 5A Buffalo NY), yielding nine building–climate configurations each released in fault-free and faulted variants. Each testbed comprises a Brick Schema v1.3 knowledge graph (zone topology, HVAC equipment hierarchy, and sensing relationships), an hourly sensor time-series store bound to the graph through shared timeseries identifiers, and an auto-gradable question-answer benchmark whose answer key is known by construction from the simulation. Sensor time-series were generated by BuildStream, which drives EnergyPlus simulations and injects faults as physics inputs rather than post-hoc perturbations, so each fault propagates through the building's control loop as a real fault would. Faulted variants each carry three injected faults (thermostat offset and outdoor-air-temperature offset) with ground-truth labels recording type, onset, magnitude, and affected KG entity. The corpus comprises 1,050 KG-linked sensor streams, 9.2 million hourly readings, 27 labeled fault instances, and 249 auto-gradable QA pairs spanning operational, diagnostic, and structural capability classes.