Machine learning has become a vital part of analysis in modern neutrino astronomy, and many
recent discoveries would not be possible without it. This approach, however, is limited by the
quality of available training data. Located at the South Pole, the IceCube Neutrino Observatory is
a neutrino detector sensitive to astrophysical neutrinos from GeV to PeV energies, with ongoing
efforts to push the sensitivity down to 100 MeV for neutrinos from transient events. IceCube is
dominated by massive backgrounds, detecting more than 10 billion atmospheric muons for each
astrophysical neutrino, and machine learning is a powerful tool to reduce this large background
rate. However, undetected outliers in labelled training data negatively affect the final performance
of machine learning algorithms. Citizen scientists can help to quantify and qualify outliers in
IceCube data to improve the detection of such outliers. In this contribution, we present the
ongoing efforts of utilising citizen science to improve a machine-learning-based event selection
targeting sub-GeV astrophysical neutrinos.