Cleaned up signatures from the dataset, new captions and new training parameters.