Top panel: simple smoothing of the training redshift distribution using SciPy’s KDE function. Bottom panel: galaxies from the highest-density regions in each of the large spikes protruding above the smoothed distribution were removed from the initial sample, resulting in the green histogram, which was then smoothed using KDE (dashed line). The hashed pink spikes show the number and redshift location of the discarded galaxies (∼ 14% in total at i<25).