Very few large-scale music research datasets are publicly available. There is an increasing need for such datasets, because the shift from physical to digital distribution in the music industry has given the listener access to a large body of music, which needs to be cataloged efficiently and be easily browsable. Additionally, deep learning and feature learning techniques are becoming increasingly popular for music information retrieval applications, and they typically require large amounts of training data to work well. In this paper, we propose to exploit an available large-scale music dataset, the Million Song Dataset (MSD), for classification tasks on other datasets, by reusing models trained on the MSD for feature extraction. This transfer learning approach, which we refer to as supervised pre-training, was previously shown to be very effective for computer vision problems. We show that features learned from MSD audio fragments in a supervised manner, using tag labels and user listening data, consistently outperform features learned in an unsupervised manner in this setting, provided that the learned feature extractor is of limited complexity. We evaluate our approach on the GTZAN, 1517-Artists, Unique and Magnatagatune datasets.