Skip to content

Dataset · google-research-datasets

NQ-Open open-domain question answering benchmark

google-research-datasets/nq_open

Classic English benchmark for testing model fact-answering capability without source context.

What's inside

The dataset is based on Natural Questions and contains 10 to 100 thousand question and correct answer pairs in English. All answers can be found in English Wikipedia, but models must answer independently during testing (Open-Domain QA task).

What you can do with it

  • Evaluating factual knowledge and hallucination levels of language models
  • Training and testing RAG systems with Wikipedia search
Tasks
Question answering
Languages
English
Size
10K–100K rows
Downloads this month
34.4K
Likes
38

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("google-research-datasets/nq_open")