This paper proposes the ACT data pipeline, which reduces human annotation costs by using MLLMs as annotators and error detectors, and provides a theoretical analysis to ensure effective downstream training.