X-ray spectroscopy is used to investigate the elemental composition and chemical states of materials through an energy spectrum, a histogram representing the energies of emitted photons. However, in high input-count-rate scenarios, the pile-up effect occurs, distorting the energy spectrum and affecting measurement accuracy. Currently, deep learning methods have been increasingly applied to pile-up correction. Nevertheless, there is a lack of open datasets for improving and measuring the performance of various algorithms. In this paper, the first open nuclear pulse dataset was constructed using the Allpix Squared [12] simulator, which is a well-recognized simulation framework developed by CERN (European Council for Nuclear Research). The dataset contains nuclear pulse data for twelve representative chemical elements. Detailed and diverse annotations are provided to support various deep-learning models. In addition, X-ray fluorescence measurements were conducted at the Shanghai Synchrotron Radiation Facility (SSRF), where the distribution of simulated data was compared with experimental data. Finally, deep learning-based pile-up correction was performed on this dataset, demonstrating its potential for training and evaluating deep learning algorithms.