This work presents an extended experimental database characterizing the transition process induced by flow separation on a flat plate with a variable adverse pressure gradient, and demonstrates its usability for tuning algebraic models of the Reynolds stress tensor by means of machine learning techniques. The database collects the 2D components of the Reynolds stress tensor for about 90 flow cases characterized by different flow Reynolds numbers, freestream turbulence intensities and adverse pressure gradients. In the range of variation chosen for the aforementioned parameters, bursting process occurs, with the consequent formation of separation bubbles of both short and long types. Sparse Bayesian Learning has been exploited to tune two algebraic turbulence models, based on the Pope’s expansion, for short and long bubbles. The models have been formulated in terms of the invariants of the strain and rotation tensors with the addition of different physical-based flow features to account for the strong anisotropy that has been observed in the transitional region. The capability of the tuned relations in reproducing the anisotropy tensor along the separated shear layer is here shown by predicting flow cases that did not participate in the tuning process. The paper aims at providing a thorough database for future researches in the field and to improve turbulence models and RANS capabilities in the simulation of separated flows.