SimpleStories: A Better Synthetic Dataset and Tiny Models for Interpretability — AI Alignment Forum