This code is used for training an RNN on large amounts of coherent text data, where each 'article' is also too large to train on in one run.
The algorithm usees characters as its basis, rather than a corpus, however does a 2 pass run where related characters are converted into words through an initial RNN before then being run through a word-level RNN.
The code supports saving and restoring from a model in the training function. Models (saved through tf.train.Saver) are placed by default into ./save