Full dataset with actual image files uploaded (no bytes, no path issue)

Full raw dataset | 3 columns: image, label, modality | No train/test split