SQuAD, the Stanford Question Answering Dataset, tests whether a model can read a Wikipedia passage and answer questions about it.
In plain words
Crowd workers read Wikipedia articles and wrote questions about them. For each question, the answer is a stretch of text copied from the passage, so a model must find the right span. Version 1.1 has over 100,000 questions. Version 2.0 adds more than 50,000 questions that cannot be answered from the passage, so a good model must also know when to say nothing.
A quick example
Imagine a passage about a river and the question "Where does it start?". The model must point to the exact words in the passage that answer it. In SQuAD 2.0, a question might ask something the passage never mentions, and the right move is to give no answer.
Where you'll meet it
- Benchmarking reading comprehension in language models.
- Training extractive question answering systems.
Related
Sources
- PaperSQuAD: 100,000+ Questions for Machine Comprehension of Text, Rajpurkar et al. (arXiv) · read 29 Sept 2026
- DocsThe Stanford Question Answering Dataset, Stanford NLP (SQuAD explorer) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.