This dataset is required to train the models in the CRISM ML toolbox [1]. In the project, we demonstrate the utility of machine learning in two essential CRISM analysis tasks: nonlinear noise removal and mineral classification. We train a hierarchical Bayesian model for estimating distributions of spectral patterns on pixel-scale training data collected from dozens of well-characterized CRISM images. The following files are included: CRISM_bland_unratioed.mat: unratioed training spectra for bland pixels. CRISM_labeled_pixels_ratioed.mat: ratioed training spectra for mineral classes. CRISM_labeled_pixel_patterns.pdf: visualization of the training segmentation maps and average spectra. The training spectra are in Matlab v7.3 (and newer) format. To load them in Python, use the mat73 library, because scipy doesn't support the format. The bland unratioed spectra have the following variables: Name Size Description pixspec 337 617 × 350 Unratioed spectra im_names 340 List of CRISM image names, mapping them to numerical IDs pixims 337 617 Numerical ID of the image the spectrum is from pixcrds 337 617 × 2 (x,y) coordinates of the points in the original image The labeled ratioed pixels have the following variables: Name Size Description pixspec 592 413 × 350 Ratioed spectra pixlabs 592 413 Mineral labels im_names 77 List of CRISM image names, mapping them to numerical IDs pixims 592 413 Numerical ID of the image the spectrum is from pixpat 592 413 ID of the connected patch in the image the pixel belongs to pixcrds 592 413 × 2 (x,y) coordinates of pixels in their respective image Citation (cite this paper when using the data): Plebani, E., Ehlmann, B. L., Leask, E. K., Fox, V. K., & Dundar, M. M. (2022). A machine learning toolkit for CRISM image analysis. Icarus, 376, 114849.