natural_questions

  • Description:

The NQ corpus contains questions from real users, and it requires QA systems to read and comprehend an entire Wikipedia article that may or may not contain the answer to the question. The inclusion of real user questions, and the requirement that solutions should read an entire page to find the answer, cause NQ to be a more realistic and challenging task than prior QA datasets.

Split Examples
'train' 307,373
'validation' 7,830
@article{47761,
title = {Natural Questions: a Benchmark for Question Answering Research},
author = {Tom Kwiatkowski and Jennimaria Palomaki and Olivia Redfield and Michael Collins and Ankur Parikh and Chris Alberti and Danielle Epstein and Illia Polosukhin and Matthew Kelcey and Jacob Devlin and Kenton Lee and Kristina N. Toutanova and Llion Jones and Ming-Wei Chang and Andrew Dai and Jakob Uszkoreit and Quoc Le and Slav Petrov},
year = {2019},
journal = {Transactions of the Association of Computational Linguistics}
}

natural_questions/default (default config)

  • Config description: Default natural_questions config

  • Dataset size: 90.26 GiB

  • Feature structure:

FeaturesDict({
    'annotations': Sequence({
        'id': string,
        'long_answer': FeaturesDict({
            'end_byte': int64,
            'end_token': int64,
            'start_byte': int64,
            'start_token': int64,
        }),
        'short_answers': Sequence({
            'end_byte': int64,
            'end_token': int64,
            'start_byte': int64,
            'start_token': int64,
            'text': Text(shape=(), dtype=string),
        }),
        'yes_no_answer': ClassLabel(shape=(), dtype=int64, num_classes=2),
    }),
    'document': FeaturesDict({
        'html': Text(shape=(), dtype=string),
        'title': Text(shape=(), dtype=string),
        'tokens': Sequence({
            'is_html': bool,
            'token': Text(shape=(), dtype=string),
        }),
        'url': Text(shape=(), dtype=string),
    }),
    'id': string,
    'question': FeaturesDict({
        'text': Text(shape=(), dtype=string),
        'tokens': Sequence(string),
    }),
})
  • Feature documentation:
Feature Class Shape Dtype Description
FeaturesDict
annotations Sequence
annotations/id Tensor string
annotations/long_answer FeaturesDict
annotations/long_answer/end_byte Tensor int64
annotations/long_answer/end_token Tensor int64
annotations/long_answer/start_byte Tensor int64
annotations/long_answer/start_token Tensor int64
annotations/short_answers Sequence
annotations/short_answers/end_byte Tensor int64
annotations/short_answers/end_token Tensor int64
annotations/short_answers/start_byte Tensor int64
annotations/short_answers/start_token Tensor int64
annotations/short_answers/text Text string
annotations/yes_no_answer ClassLabel int64
document FeaturesDict
document/html Text string
document/title Text string
document/tokens Sequence
document/tokens/is_html Tensor bool
document/tokens/token Text string
document/url Text string
id Tensor string
question FeaturesDict
question/text Text string
question/tokens Sequence(Tensor) (None,) string

natural_questions/longt5

  • Config description: natural_questions preprocessed as in the longT5 benchmark

  • Dataset size: 8.91 GiB

  • Feature structure:

FeaturesDict({
    'all_answers': Sequence(Text(shape=(), dtype=string)),
    'answer': Text(shape=(), dtype=string),
    'context': Text(shape=(), dtype=string),
    'id': Text(shape=(), dtype=string),
    'question': Text(shape=(), dtype=string),
    'title': Text(shape=(), dtype=string),
})
  • Feature documentation:
Feature Class Shape Dtype Description
FeaturesDict
all_answers Sequence(Text) (None,) string
answer Text string
context Text string
id Text string
question Text string
title Text string