I think there is a small bug in dataset.py that affects the building of the Mention Type one-hot vectors of antecedent mentions in the pair features during training. Due to the use of slicing by a colon in the first dimension, the assignment is made on the full columns referred by the index in the 1-D array ant_features_raw[:, 0], which contains the mention type of the antecedent mentions expressed as integer. The expected behaviour I think was to put at 1 a single bit only, indexed by the 1-D array, for each row/antecedent mention, as it's done for the main mention.

This causes a mismatch between the training features and the inference ones: in neuralcoref.pyx, the mention type is correctly encoded as a one-hot vector for each mention, and then copied in the pair features for the antecedent mentions.

This is a simple example with numpy comparing actual vs expected results: