The dataset is based on a heterogeneous polymetallic mineral sample originating from the W–Au–Pb–Zn–Ag (Sb–Ba) district of Tighza, Morocco, a well-documented mining area characterized by complex hydrothermal mineralization. The sample exhibits strong chemical and mineralogical heterogeneity at the microscale, including quartz, galena, ankerite, and aluminosilicates, and was selected as a challenging benchmark for Raman spectral analysis and Noise2Noise-based denoising.This dataset contains raw one-dimensional Raman spectra acquired from a heterogeneous polymetallic mineral sample at multiple integration times (5, 10, 25, 50, and 100 ms). The data are provided in HDF5 (.h5) format and include repeated acquisitions at each exposure time, organized to enable the construction of dedicated training datasets for each integration time. All spectra are supplied prior to any denoising or machine-learning-based processing. The repeated measurements capture realistic noise statistics typical of high-throughput Raman mapping and form the experimental basis for training and benchmarking Noise2Noise denoising approaches, as reported in the associated study.