Dataset · google-research-datasets
User questions based on Wikipedia articles
google-research-datasets/natural_questions
English QA benchmark: real user questions requiring the reading of a Wikipedia article to answer.
What's inside
A corpus of real user questions in English. To answer, a QA system must read an entire Wikipedia article, which may or may not contain the answer, making the task more realistic and difficult than many previous QA datasets. Volume ranges from 10k to 100k examples with text modality.
What you can do with it
- Evaluating QA systems on real questions
- Testing retrieval pipelines for RAG
- Fine-tuning models to answer based on full Wikipedia articles
- Tasks
- Question answering
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 40.9K
- Likes
- 133
Models trained on it
- vllm-sr/Vela-2.0-0.8BZero shot classification 1.9K
- vllm-sr/Vela-2.0-4BZero shot classification 249
- vllm-sr/Vela-2.0-9BZero shot classification 245
- parsecai/ragzillaText classification 56
- syedkhalid0/flan-t5-small-fidText generation 35
- seamon67/Vela-2.0-Omni-4BZero shot classification 31
- AnjanSB/Qwen2.5-3B-Instruct-NQ-RAG-DPO-LoRAText generation 16
- vsearch/svdr-nq 13
How to load it
from datasets import load_dataset
ds = load_dataset("google-research-datasets/natural_questions")