Third trial dataset for the SemEval 2019 Task 4: Hyperpartisan News Detection. The dataset contains 1 million articles. It is split in training (200,000 left, 400,000 least, 200,000 right) and validation (50,000 left, 100,000 least, 50,000 right), where no publisher that occurs in the training set also occurs in the validation set. All articles are labeled by the overall bias of the publisher as provided by BuzzFeed journalists or MediaBiasFactCheck.com. The trial data is not fully cleaned. Due to some encoding error, some characters are replaced by question marks. However, all files are already fully compatible with the XML schema files.