DROP Explorer

This is a data explorer for the dataset DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs.

Available for immediate use at http://eladsegal.github.io/DROP-explorer.

• Dataset files are available here.
• Prediction files examples are available here.
  For DROP, use the standardized dataset file here for correct alignment.
  For Quoref, use the DROPified dataset file here.

Things to know

• Clicking on a question row will highlight a gold answer and the predicted answer.
• Multi-span answers are sorted
• The first answer displayed is the one in "answer", the rest are from "validated_answers" and only distinct answers are displayed.
• Be aware that heads with some names are handled in a special way:
  • Predictions from "counting" head won't be highlighted
  • Predictions from "arithmetic" head are expected to have a member named "numbers" in "answer" which is an array.

Expected Predictions File Format

The expected predictions file format is JSONL, where each line is a JSON object that is the output_dict of an instance prediction. The following members will be used and are required unless mentioned otherwise (they don't have to be correct, but just have the correct type):

• passage_id: string
• query_id: string
• answer: A JSON object with the following members:
  • value: The final prediction - A string or an array of strings
  • spans (optional): An array of arrays of the form ["p" (passage) / "q" (question), start_index, end_index (exclusive)] used to make spans that the model used for prediction bold.
  • numbers (required and used only when the head is "arithmetic"): An array of objects of the form {"value": number, "sign": -1/0/-1} to construct the arithmetic expression used to arrive at the answer
• predicted_ability: The name of the head used for prediction
• maximizing_ground_truth: The answer for which the highest EM and F1 scores were calculated, in the same format of an answer in the dataset.
• em (optional): The EM score calculated, a number
• f1 (optional): The F1 score calculated, a number
• max_passage_length (optional): The length of the passage that was considered for the model prediction, used to show which parts of the passage were truncated