Towards Scalable Meta-Learning of near-optimal Interpretable Models via Synthetic Model Generations