The metrics below show that every hyperparameter of the process is important, even dataset size and epochs number. Therefore the modelling process requires a lot of attention.
The overfitting can be verified by comparing training and validation losses function. It happens due to too few training data and too many epochs.

nn_plot(hist[-1], 200, loss=True, merge=False)

The training accuracy reaches 100%, which should be avoided, since an accuracy of 100% usually means overfitting. The validation accuracy remains still at a lower value. While the overfitting is highlited in the loss function, it requires some interpretation in the accuracy graphic.

nn_plot(hist[-1], 200, acc=True)

One way to reduce overfitting, even with a small dataset available, is by applying regularization. We will mantain the small training dataset and high number of epochs, on purpose, to compare the results applying regularization to same conditions.
The loss function is penalized now by a $\lambda$ factor. We will apply L2 Regularization, with $\lambda = 0.01$, its default value in keras.

model.append(tf.keras.models.Sequential([
    tf.keras.layers.Dense(30, kernel_regularizer=tf.keras.regularizers.l2(0.01),
                          activation=tf.nn.sigmoid, input_shape=(784,), name='hidden_1_layer'),
    tf.keras.layers.Dense(10, activation='softmax', name='output_layer')
]))

nn_compile(model[-1], params[0])