New training algorithm for linear classification SVMs that can be much faster than SVMlight for large datasets. It also lets you direcly optimize multivariate performance measures like F1-Score, ROC-Area, and the Precision/Recall Break-Even Point.