Path explosion is a crucial problem in symbolic execution that leads to poor performance in symbolic execution engines and hinders the widespread adoption of respective tools. A path selector is a component of a symbolic machine designed to address the path explosion problem. AI-powered path selectors have gained attention, but many challenges regarding the training process, feature selection, and information representation remain. We propose PySymGym, a framework for training AI-powered path selectors through typical supervised learning, which includes language-independent, graph-based data representation, a training protocol that minimizes manual dataset preparation, and supportive infrastructure. Evaluation of the proposed solution shows that it enables training models comparable to searchers based on manually developed heuristics: providing close coverage percentage at comparable analysis time (with the same timeout), and allowing the system to generate fewer tests.