SmolTulu: Higher Learning Rate to Batch Size Ratios Can Lead to Better Reasoning in SLMs