The data comes from the yelp review challenge, which is a document classification challenge: how many stars did each review get?

It has been sampled and processed as follows:

• words.txt -> vocabulary, one word per line
• documents.txt -> dataset, one document per line. Sentences separated by |&|, words already tokenized
• labels.txt -> int label for each document
• tf_multi_sequence_dataset.py -> standalone code exhibiting how to store sequences of sequences into a tf.data.Dataset

python: 3.6.4 tensorflow: 1.8

$ python tf_multi_sequence_dataset.py
(32, 28, 55)
(32, 57, 112)
True
True