Raman fiber amplifier have been widely studied because of their low noise and wide gain bandwidth. Aiming at the complex time-consuming and low efficiency of traditional numerical solution methods, this paper proposes a new inverse design scheme for Raman fiber amplifiers combining convolutional neural network and support vector regression algorithms to achieve ultra-fine, dynamic and arbitrary Raman Amplify gain spectrum shaping. First, the performance of the model under different parameter settings is studied through the cooperative search algorithm to determine the best learning model, so that the model can more accurately reflect the mapping relationship between the Raman gain spectrum and the pumping parameters. The ability of the model to shape the target flat and arbitrary gain spectrum is analyzed. The simulation results show that the proposed model achieves a maximum gain flatness of 0.1962 dB in the design of a flat target gain Raman fiber amplifier in the C L band. For the target of arbitrary gain Raman fiber amplifier design, the maximum root mean square error(RMSE) is only 0.13 dB, and the average design time is only 4.511?10-5s . To further measure of arbitrary gain spectrum optimization, the probability density functions(PDF) and the Cumulative Density Function(CDF) of RMSE and Max Error are proposed. The presented method could extend the application of data-driven methods in the field of optical amplification, and provides a new method and idea for the flexible and rapid design of Raman fiber amplifiers in the future.