Qasper — Questions/Answers on scientific papers
Qasper is a scientific QA corpus based on full-text NLP articles, with manual annotation of answers and evidence within the document. Ideal for long reading or literature review models.
5,049 questions on 1,585 articles, structured JSON (text, questions, answers, annotations)
CC-BY 4.0
Description
Qasper is a dataset designed for the question and answer on scientific articles in NLP. Each example is based on a complete scientific article, with one or more questions formulated from the title and abstract, and then answered by reading the entire text. The answers are accompanied by evidence extracted from the body of the text, making them an excellent data set for training and evaluating “long read” or “reasoning with evidence” models.
What is this dataset for?
- Train models that can read an entire scientific article to answer specific questions
- Test Longformer or BigBird architectures on evidence selection tasks
- Develop synthesis or assisted scientific QA systems
Can it be enriched or improved?
Yes. It is possible to extend this corpus to other scientific fields (beyond NLP), to translate the questions to create a multilingual version, or to reinforce the evidence annotations with confidence scores or reformulations. The addition of additional annotations (level of difficulty, nature of the question, etc.) would also allow for more advanced uses.
🔎 In summary
🧠 Recommended for
- NLP engineers
- Researchers
- Autoplay engine developers
🔧 Compatible tools
- Longformer
- BigBird
- Hugging Face Transformers
- OpenQA pipelines
💡 Tip
To maximize impact, combine Qasper with specialized models for scientific summary or semantic exploration.
Frequently Asked Questions
Is Qasper limited to NLP items?
Yes, the current version of the dataset focuses only on automatic language processing (NLP) articles.
Can a model be trained to select the most relevant passages with Qasper?
Absolutely, Qasper provides obvious annotations that are ideal for the task of automatically selecting passages.
Is this dataset suitable for long models such as Longformer or extended GPT?
Yes, the full-text format with multiple paragraphs makes it ideal for long reading architectures.
