The held-out scores reported in the Readme IS the unseen dataset.
These are not successful prediction rate per char though.
These are not successful prediction rate per char though.