New Research Demonstrates Enormous Efficiency Gains in Language Model Training Through Modular Data Ablation