Learning rate is the most critical hyperparameter. Too low: model trains too slowly. Too high: oscillates / diverges. This tool sweeps 10 LR values from 0.001 to 0.5 and reports final loss + accuracy for each.