A data-driven approach to cleaning image classification datasets using embedding space visualization and UMAP dimensionality reduction to filter out bad training data and enhance machine learning model performance.