This paper presents a novel dual coordinate descent method for solving large-scale linear Support Vector Machines (SVM) with L1- and L2-loss functions. The method is designed to handle datasets with a large number of instances and features, making it particularly suitable for applications such as document classification. The proposed method is simple to implement and achieves an $\epsilon$-accurate solution in $O(\log(1/\epsilon))$ iterations. Experimental results show that the method outperforms state-of-the-art solvers like Pegasos, TRON, SVMperf, and a primal coordinate descent implementation in terms of speed. The paper also discusses the implementation details, including random permutation of sub-problems and shrinking techniques, which further enhance the efficiency of the method. Additionally, the paper compares the proposed method with other linear SVM methods and provides insights into its performance and convergence properties.This paper presents a novel dual coordinate descent method for solving large-scale linear Support Vector Machines (SVM) with L1- and L2-loss functions. The method is designed to handle datasets with a large number of instances and features, making it particularly suitable for applications such as document classification. The proposed method is simple to implement and achieves an $\epsilon$-accurate solution in $O(\log(1/\epsilon))$ iterations. Experimental results show that the method outperforms state-of-the-art solvers like Pegasos, TRON, SVMperf, and a primal coordinate descent implementation in terms of speed. The paper also discusses the implementation details, including random permutation of sub-problems and shrinking techniques, which further enhance the efficiency of the method. Additionally, the paper compares the proposed method with other linear SVM methods and provides insights into its performance and convergence properties.