ROCStories (prompt / continuation)

A reformatted version of the ROCStories corpus, suitable for open-ended story-generation exercises.

Each example is a 5-sentence ROCStory, split into:

field	description
prompt	the first sentence of the story
continuation	the remaining four sentences
text	the full (unmodified) 5-sentence story

Splits

split	rows
train	70,676
validation	7,852
test	19,633

The train / validation split is a 90/10 shuffle (seed = 42) of the train split from mintujupally/ROCStories; test is mintujupally's test split unchanged.

Provenance

This dataset was reconstructed to replace Ximing/ROCStories, which is no longer available on the Hub, as the data source for the CSED503 / CSE447 Natural Language Generation assignment.

The underlying stories come from the ROC Stories 2016 and 2017 releases (Mostafazadeh et al.). Please cite the original corpus when using this data:

@inproceedings{mostafazadeh2016corpus,
  title     = {A Corpus and Cloze Evaluation for Deeper Understanding of Commonsense Stories},
  author    = {Mostafazadeh, Nasrin and Chambers, Nathanael and He, Xiaodong and Parikh, Devi and Batra, Dhruv and Vanderwende, Lucy and Kohli, Pushmeet and Allen, James},
  booktitle = {Proceedings of NAACL-HLT},
  year      = {2016}
}