What Should Baby Models Read? Exploring Sample-Efficient Data Composition on Model Performance