The dataset contains three files:
* training.csv: 119 original sentences each with four human simplifications and ten sentence fusion sentences all scored for adequacy and simplicity used for training the ranker.
* test.csv: 200 original sentences each with four human simplifications all scored for adequacy, simplicity, and fluency.
* development.csv: 119 original sentences with two sets of ten sentence fusion sentences from the first two iteratins of training the ranking, all scored with adequacy and simplicity.  The final results from the paper only trained on the training data, but this data may be useful for other purposes.

The four human simplifications were generated using Amazon Mechanical Turk.

The simplicity and adequacy scores were as well and are the result of averaging scores from three workers.
