How many epochs should all the data be trained on after training with validation finds when validation and training diverge?

One uses and train/test split to use their training data to get an idea of how many epochs to train with. If the validation accuracy starts going down while the training accuracy is still going up, this would show a sign of overfitting, so one should probably stop the number epochs around there. But should training with all the data take longer to overfit? If so, should one add a few epochs? And if so for that, is there a good rule of thumb for how many? Maybe training different amounts of data could give an idea of what the differences are for number of epochs for the same amount of accuracy, even though they might not reach the same maximum amount. Any ideas on this?