The paper considers the problem of adapting the model to new data with a large amount of information. We propose to build a more complex model using the parameters of a simple one. We take into account not only the accuracy of the prediction on the original samples but also the adaptability to new data and the robustness of the obtained solution. The work is devoted to developing the method that allows adapting the pre-trained model to a more heterogeneous dataset. In the computational experiment, we analyse the quality of predictions and model robustness on Fashion-MNIST dataset.