We introduce the LS3DS dataset, a novel collection of large 3D CAD models, meshes, and point clouds of industrial sites. Additionally, the proposed dataset provides a processing pipeline to generate synthetic meshes and point clouds from the CAD models. The open-source pipeline makes the proposed dataset easily applicable to different scenarios, such as the construction industry. The ground truth values can be effortlessly generated to tackle problems such as geometric primitive fitting, a meaningful challenge strongly related to segmentation and shape fitting problems. Furthermore, we present a benchmark addressing the problem of large-scale point cloud geometric primitive fitting. We adapted state-of-the-art deep learning-based methods for the benchmark to process large-scale point clouds. We compared them to a baseline classical approach, which shows challenges in complex, large-scale industrial environments defined by dense and varied geometric distributions. Our paper demonstrates the meaningful contribution of the proposed dataset with a case study presented in the benchmark, opening opportunities for dealing with relevant problems of 3D geometric understanding using learning approaches. The generation pipeline, LS3DS dataset and the weights of the models trained in the benchmark are openly available to use (LS3DS Repository: https://github.com/igormaurell/LS3DS ).