The dataset is based on a debate.org crawl. It is restricted to a subset of four out of the total 23 categories -- politics, society, economics and science -- and contains additional annotations. 3 human annotators familiar with linguistics segmented these documents and labeled them as being of medium or low quality, to exclude low quality documents. The annotators were then asked to indicate the beginning of each new argument and to label argumentative sentences summarizing the aspects of the post as conclusion and outside of argumentation. In this way, we obtained a ground truth of labeled arguments on a sentence level (Krippendorff's alpha=0.24 based on 20 documents and three annotators).