Notes. Classification is done by assigning the class with the largest posterior probability. Performance is given in terms of completeness (compl.) and purity, for each classifier and for each class. Purities have been adjusted to reflect the class prior (given in Table 2). Results on the ‘binary’ class are largely meaningless due to the incongruity of the class definitions in the training and validation data sets. These results reflect performance for sources drawn at random from the entire Gaia data set, in particular for all magnitudes and latitudes. The final two columns labelled ‘Spec&Allos’ refer to samples obtained by requiring a probability larger than 0.5 from both Specmod and Allosmod for a given class: this is identical to classlabel_dsc_joint in the qso_candidates and galaxy_candidates tables. The bottom two rows refer to extragalactic sources at higher Galactic latitudes (|b|> 11.54°), where the prior is more favourable for detecting quasars and galaxies. These are conservative estimates, accounting only for reduced numbers of stars, not the better visibility of extragalactic objects on account of less interstellar extinction and source confusion.