zlacker

I assume lock and dependency files are in the training data, so predicting version number tokens have high probabilities associated with them.