Skip to content

Dataset · UCLNLP

Adversarial questions for testing text comprehension

UCLNLP/adversarial_qa

Three reading comprehension question sets created to identify weaknesses in language models.

What's inside

An English text dataset for Reading Comprehension tasks, collected with neural network participation in a labeling loop (BiDAF, BERTLarge, and RoBERTaLarge). It consists of three separate parts of 12,000 examples each (10,000 for training, 1,000 for validation, and 1,000 for testing). The questions are selected to confuse baseline models when searching for an answer in the text.

What you can do with it

  • Testing machine reading models' robustness to complex and tricky questions
  • Fine-tuning QA systems to improve the accuracy of extracting answers from text
  • Comparing transformer architectures on adversarial examples
Tasks
Question answering
Languages
English
Size
10K–100K rows
Downloads this month
5.2K
Likes
44

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("UCLNLP/adversarial_qa")