AI Redefining Efficiency: Beyond Compute-Optimal Training to Predict Language Model Performance on Downstream Tasks