The dynamics modeling problem of flight vehicles is studied in this paper. A new data-driven modeling approach is presented. This approach consists of a system item fitting model based on a neural network and a kind of dynamic model structure based on state-space expressions. Given the traditional data-driven dynamics modeling process of flight vehicles, this paper proposes a solution to the problem of model divergence caused by error accumulation and the problem that data-driven models cannot be combined with classical control strategies. The neural network is first trained with the system term in the state-space expression of the attitude dynamics equation as the output data to improve the modeling accuracy and suppress the divergence of the model. Then, the data-driven attitude dynamics model of the flight vehicle is indirectly constructed according to the state-space form. Simulation show the effectiveness of the proposed data-driven modeling approach.