Could you share the training loop and model details as well? Little hard to tell based purely off the data.