Social cohesion and polarization can be studied with society-wide networks, however such data are not easy to collect. The alternatives—stylized networks often used in agent-based models (ABM) and small-scale empirical networks—might misrepresent societal structure at the macro-level, in terms of both topological properties and differences in network positions of social groups. We demonstrate how more realistic society-level networks can be generated from ego-network data that can be reasonably collected even for large populations. The empirical data we use come from a large-scale survey on a representative national sample about individuals’ core discussion ties, collected by the Spanish Institute for National Statistics (INE). In this paper, we describe our methods and the properties of the resulting networks. We find that large-scale core discussion networks do not have small-world properties, resembling population-level familial networks rather than friendship networks. At the same time are characterized by assortativity on dimensions such as age, gender, working status and political views. We also demonstrate how model-based simulations of complete networks can be further analyzed to make inferences that go beyond personal network analysis, such as centrality and connectivity. Together with this paper, we release the simulated full networks for secondary use by ABM modelers and wider research community interested in studying cohesion in realistic networks.