BOB12311/natural-questions-slim-short-answer

Natural Questions Slim Short Answer

This is a slim, flattened derived version of google-research-datasets/natural_questions for short-answer question answering experiments.

The conversion keeps examples with extractable short answers and removes the original document HTML, token-level document spans, long answer candidates, and yes/no-only examples. Each record is a simple question-answer pair. It is intended for lightweight QA prompting and evaluation, not as a full replacement for the original Natural Questions dataset.

Dataset Structure

Each JSONL record contains:

• id: original Natural Questions example id as a string
• question: question text
• answer: first deduplicated short answer

Example:

{
  "id": "4549465242785278785",
  "question": "when is the last episode of season 8 of the walking dead",
  "answer": "March 18 , 2018"
}

Citation

@misc{sun2026pathmatters,
      title={The Path Matters: Learning a Token-Commitment Policy for Diffusion Language Models}, 
      author={Bohang Sun and Max Zhu and Francesco Caso and Jindong Gu and Junchi Yu and Philip Torr and Pietro Liò and Jialin Yu},
      year={2026},
      eprint={2605.24697},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.24697}, 
}

@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}
}

Splits

Split	Source examples	Kept examples	Keep ratio
train	307,373	106,926	34.7871%
validation	7,830	4,289	54.7765%

The build produced 294 JSONL shard files under shards/.

Filtering

The conversion script:

• normalizes whitespace and HTML entities in questions and answers
• extracts non-HTML token spans for short answers
• drops examples without short answers
• drops yes/no-only examples
• deduplicates short answers while preserving order
• keeps the first deduplicated short answer as answer

The source conversion script in the Remasker repository is prepare/build_natural_questions_slim.py.

Loading

from datasets import load_dataset

dataset = load_dataset("BOB12311/natural-questions-slim-short-answer")
train = dataset["train"]
validation = dataset["validation"]

Source and License

This dataset is derived from Natural Questions as distributed by google-research-datasets/natural_questions on Hugging Face.

The source dataset is licensed under Creative Commons Attribution-ShareAlike 3.0 Unported (cc-by-sa-3.0). This derived dataset is distributed under the same license. Users should cite and comply with the terms of the original dataset.

Source dataset: https://huggingface.co/datasets/google-research-datasets/natural_questions

Intended Use

This dataset is intended for research on short-answer QA prompting, reranking, generation control, and lightweight evaluation. Because long documents and full annotation structures are removed, it should not be used for experiments that need Natural Questions long-answer evidence, document retrieval, or span-level HTML/token alignment.