In this paper, a distributed data-driven frequency control scheme is proposed for the off-grid networked microgrid system (NMG). In the proposed scheme, the frequency is regulated by coordinately adjusting the active power generation provided by controllable generators and the active power demand provided by voltage sensitive loads, where the active power consumption of voltage sensitive loads is specially modified by adjusting the load operating voltages via voltage regulation. The whole control problem is solved with the feedback-based gradient method. To achieve better scalability and privacy, the overall control scheme is implemented in a distributed manner by employing an approximated gradient that can be calculated by limited communication between neighboring MGs. To further enable the controller to adapt to real-time system conditions, the required system model is updated online via the partial least square method in a recursive manner, which renders the proposed controller model-free. Numerical simulations involving detailed nonlinear NMG models illustrate the key benefits of the proposed method, including (i) eliminating the need for an accurate offline system model, (ii) reducing the need of controllable generation, and (iii) being robust against communication delays.