Adapted from http://www.cs.cmu.edu/~glai1/data/race/, which is the source of the original RACE dataset.

Each passage is a JSON file. The JSON file contains following fields:

* Fields from original RACE data:
  - article: A string, which is the passage.
  - questions: A string list. Each string is a query. We have two types of questions. The first one is an interrogative sentence. The second has a placeholder, which is represented by _.
  - options: A list of the options list. Each options list contains strings, which are the candidate option.
  - answers: A list contains the golden label of each query.
  - id: Each passage has a unique id in this dataset.
  NOTE: the original race dataset contains precisely 4 choices per question. This dataset contains 4 or more choices per question.

* Fields added as part of this paper:
  - questionGroupIds: A list indicating which variant group the question belongs to. Questions in the same variant group are meant to ask for essentially the same information. Systems can be scored in an all-or-nothing manner on each variant group.
  - questionType: a tag indicating whether the question is spatio-temporal (ST) or causal/motivational (CM).