This models were trained with two classes: Patacón-True and Patacón-False. Anything that could create a bias towards a feature that could confuse the model was emphasized on specific data fed to the False class. A greater number doesn't correlate with a better performance, it's a cronological order, but later versions are (sometimes) better than the early ones.
The most recent version is here: https://huggingface.co/spaces/frncscp/Patacotron