We examine the impact of both training and test data selection in machine learning applications for abstract argumentation, in terms of prediction accuracy and generalizability. For that, we first review previous studies from a data-centric perspective and conduct some experiments to back up our analysis. We further present a novel algorithm to generate particularly challenging argumentation frameworks wrt. the task of deciding skeptical acceptability under preferred semantics. Moreover, we investigate graph-theoretical aspects of the existing datasets and perform some experiments which show that some simple properties (such as in-degree and out-degree of an argument) are already quite strong indicators of whether or not an argument is skeptically accepted under preferred semantics.