To train general networks, the first step was assembling a large, diverse training set. In the end it comprised 4,000+ tilt series from 25+ species and 58 unique data sources, covering all common detector types and everything from purified samples to lifted-out tissue.