Electronic sports (eSports), the sport competition using video games, has become one of the most popular sporting events now. The eSports audience needs textual commentaries for deeply understanding the games and for efficiently retrieving specific games of their interest. Therefore, in this work, we set up an eSports data-to-text generation task and tackle three fundamental problems: dataset construction, model design, and evaluation metrics. We first build a data-to-text dataset containing data records and game commentaries from the a popular eSports game, League of Legends. On this new dataset, we propose a hierarchical model to address difficulty in handling long sequences of inputs and outputs with an encoder-decoder model. The hierarchical model sets multi-level encoders for the input data. Besides, we organize and design a new set of evaluation metrics including three aspects to meet this task’s goal. Experimental results on the new datasets confirm that the hierarchical structure improves the performance of the model.