Sample distribution after the object-level split by spectral type. The training, validation, and test sets follow an 8:1:1 ratio. All augmented samples from the same physical object were assigned to a single subset to avoid data leakage. Images containing M dwarfs are included as nontarget contaminant samples. They are shown only to illustrate the spectral-type composition of the input data and are not treated as an independent detection class. The plotted counts represent augmented image cutouts and should not be interpreted as the intrinsic spectral-type distribution of unique Pan-STARRS sources.