Quiz Bowl Dataset
The QANTA tossup dataset is updated annually with this year's version being referred to as "QANTA 2018". The dataset is described in our arXiv preprint. These data are useful for training systems on the QANTA shared task. You can download the dataset at the links below, or use a python script to download them.
The QANTA dataset is based on the Wikipedia dumps from 4/18/2018. Since they are no longer available at the regular dumps location we also provide a copy below. For convenience we also provide a json file which contains only Wikipedia pages for answers in the dataset.
We also provide the preprocessed datasets to help build Machine Reading Comprehension (MRC) based models. We split the questions into individual sentences. For each sentence, we provide the top-5 sentences over the whole wikipedia using TFIDF scoring.
For each of train, dev, test, there will be two 'evidence' files (rough description below):
1. *evidence.json : The evidence consists of the top 5 sentences for every sentence for every question in the form of the wiki page name, and the paragraph and sentence index in that wiki page. This also has the correct answer span in these documents (sentences) (will be an empty list if answer not present in these sentences).
2. *evidence.text.json: Same as above, except the actual sentence text at every instance is included.
Qanta QBLink Dataset
The training, development, and testing splits of Qanta sequential question-answering dataset, QBLink, are available at
The data are described in a EMNLP 2018 paper.
Qanta Adversarial Dataset
A test set (~1000 questions) that challenge both humans and computers. Created using the adversarial writing process described in our TACL paper: "Trick Me If You Can: Human-in-the-loop Generation of Adversarial Question Answering Examples". The data has the exact same format as the QANTA 2018 data posted above.
Past data released by Mohit Iyyer.
Paragraph-based Context Data
We also provide the much larger paragraph based context files to help build Machine Reading Comprehension (MRC) based models. We split the questions into individual sentences. For each sentence, we first retrieve top-10 wikipedia articles over whole wikipedia using TFIDF scoring. Then inside these articles, we retrieve top-10 paragraphs with TFIDF scoring as candidates. We use TAGME to extract all entities linked to wikipedia page for each retrieved paragraph.
In the provided json file, for each question, it has the "annotated_paras" property with the list of extracted paragraphs. And each paragraph has text("paragraph") and Tagme annotated entities("entities") with (Wikipedia title, start char index, end char index, score, Wikipedia id).