In this manuscript, we construct physics-informed neural network and improved physics-informed neural network by modifying the loss function, for predicting the dynamic behaviors of bright-bright single-peak solitons, bright-bright double-peak solitons and dark-bright single-peak solitons for the coupled Sasa-Satsuma equations, which depict the characteristics of two ultra-short pulses with the third-order dispersion, stimulated Raman scattering effects and self-steepening propagating simultaneously in birefringent or dual-mode fibers. Firstly, the physics-informed neural network, which is a standard model for managing the soliton prediction, is improved to a double-layer structure, to forecast the bright-bright single-peak solitons. When predicting the bright-bright double-peak solitons and dark-bright single-peak solitons, we find that the above model does not learn the dynamics of solitons, so we add the end-time conditions as the constraints according to the motion characteristics of dynamic solitions. At the same time, considering the complex boundary conditions of the dark solitons, we modify the boundary conditions in the loss function of improved physics-informed neural network for predicting bright-dark solitons. By capturing instantaneous plots at three different times and comparing the predicted values with the exact solutions, it shows that the improved physics-informed neural network is effective. Furthermore, we select the appropriate number of iterations according to the comparison of training error and training time to improve the accuracy of the model.