The repository https://github.com/tensorflow/models has a lot of useful functionality such as its MNIST loading module. However, I don't think any of the code is exposed as part of the public Python TensorFlow interface.
Why is it not part of the public API?
Seems like a sensible to me. These are examples where the TensorFlow library has been used in order to solve a specific problem/example.
It makes sense that the code used to solve these problems should be entirely separate from the main library/framework API, no?
ALthough, if you are specifically interested in loading the NIST dataset note that there's now an example dataset available under the keras api:
mnist = tf.keras.datasets.mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()