Is the training data available anywhere? And how accurate would you expect this method to be with more training data?