Psychological and neuroscientific research on music-evoked emotions has been limited by certain stimulus characteristics. Typically, the stimuli are either artificially manipulated, unavailable for sharing due to copyright restrictions, sparsely sampled from narrow musical genres, or skewed by experimenters’ biases. This paperpresents empirical evidence that carefully curated subsets of a large-scale open dataset are comparable to well-matched commercial music in terms of various subjective ratings, including liking and feeling moved. In addition, we demonstrate the potential of a generative music service for use in human experiments. Here, we publish an open music audio dataset, curated through music information retrieval (MIR) model predictions and human validation. We believe this dataset will be highly beneficial for empirical music research and MIR research.