Skip to content

Dataset · ucinlp

DROP: Discrete reading benchmark

ucinlp/drop

96k paragraph-based questions to test arithmetic, counting, and sorting.

What's inside

The English-language DROP benchmark consists of 96k crowdsourced and adversarially created questions. Systems must resolve references in the question, potentially to multiple places, and perform discrete operations such as addition, counting, and sorting. The data is designed for reading comprehension and question answering.

What you can do with it

  • Evaluate QA models on discrete operations
  • Fine-tune reading comprehension models
  • Compare LLMs on the DROP benchmark
Tasks
Question answering
Languages
English
Size
10K–100K rows
Downloads this month
14.6K
Likes
72

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("ucinlp/drop")