Props to the author for releasing the data. My instinct is also to immediately suspect data leakage. It's super easy for this to happen. For example the original dataset could contain multiple articles about the same event.