Non-linear soil behaviour adds complexity in accurate parameter selection for numerical modelling. One of these pa-rameters is the small-strain shear stiffness. This parameter depends strongly on the soil mass density and the shear wave velocity; the latter can be determined through in-situ tests or laboratory tests. The paper focuses on training vari-ous machine learning models to predict shear wave velocity estimates based on raw data from cone penetration test soundings. Three decision tree algorithms are considered for the analysis: XGBRegressor, HistGradientRegressor, and RandomForest. Various data preprocessing approaches are investigated, including noise removal and outlier identifica-tion, to assess their impact on the model performance. The results indicate that different data preprocessing approach-es yield significant differences in the model performances. When applied to unseen raw data from a sand site of the Norwegian GeoTest Site, the model demonstrates promising predictive capabilities and is in a good agreement with well-known correlations. This study underlines the importance of data quality and preprocessing for reliable machine learn-ing models. To enhance transparency and reproducibility, a GitHub repository with all the used files is made available online.