After some investigation, I have found the problem. The model included two Dropout layers, which are active during training but disabled during evaluation. This was affecting the final accuracy in evaluate, as the model sees all connections in use during inference. After removing the Dropout layers, the model was able to train correctly.