This paper presents a new approach to construct the training dataset for the Numerical Dispersion Mitigation network (NDM-net). The network was designed to suppress numerical error in seismic modeling results. In this approach, a small number of seismograms generated using coarse and fine girds are used to train the network, mapping the inaccurate coarse-grid data to the high-quality fine-grid data. After that, the network is applied to the entire set of seismograms, precomputed using the coarse grid to reduce the numerical error. The most time-consuming part of the suggested approach is the training dataset generation. Thus, we need to minimize the number of seismograms in the training dataset without the loss of the training quality. We suggest constructing the training dataset preserving the Hausdorff distance between the training dataset and the entire dataset. However, the level of the limiting distance may vary depending on the seismogeological model used for simulation. We illustrate that the adaptive strategy is preferred over the fixed Hausdorff distance limiting level because it allows reducing the training dataset without loss of accuracy.