iapp_wiki_qa_squad

Extractive question answering over Thai Wikipedia articles, in SQuAD format. 7,242 questions across 1,912 articles, annotated by people iApp hired for the purpose.

from datasets import load_dataset

dataset = load_dataset("iapp/iapp_wiki_qa_squad")

This works again as of the August 2026 revision. Until then it did not. The repository carried a loading script and no data, and datasets dropped script support at v3, so load_dataset failed and every datasets-server endpoint returned HTTP 500. The data now lives in the repository as parquet, and the script is gone. Nothing about the data changed — see "What changed, and what did not" below.

Source and licence

Underlying text	Thai Wikipedia, CC BY-SA
Questions and answer spans	iApp Technology, annotated 2019
Canonical source	iapp-technology/iapp-wiki-qa-dataset, squad_format/data.zip — MIT
SQuAD-format processing	@cstorm125
This dataset	MIT, see LICENSE
DOI	10.5281/zenodo.4539916

The MIT grant covers iApp's contribution — the questions, the answer spans and the assembly. It does not reach the context column, which is Wikipedia article text under CC BY-SA and stays under CC BY-SA in any redistribution. LICENSE sets this out.

Translation provenance

Not a translation. Every field is original Thai: the contexts are Thai Wikipedia articles, and the questions and answers were written in Thai by human annotators. There is no source language and no machine translation anywhere in this dataset.

This section exists because every iapp/* dataset card is required to answer the question. Answering it with "not applicable, and here is why" is the answer.

Fields

field
question_id	e.g. 0U2lA8nJQESIxbZrjZQc_000
article_id	groups questions belonging to the same article
title	article title
context	article text, the passage the answer is extracted from
question	Thai
answers	{text: [...], answer_start: [...], answer_end: [...]}, one answer per question

answer_end is an exclusive upper bound. Every question has exactly one answer; the list shape follows the SQuAD convention rather than indicating multiple answers.

All 7,242 answer spans are exact: context[answer_start:answer_end] == text holds for every row in every split, with no exceptions. validate.py checks it.

An example

{'question_id': '0U2lA8nJQESIxbZrjZQc_000',
 'article_id': '0U2lA8nJQESIxbZrjZQc',
 'title': 'สุวัฒน์ วรรณศิริกุล',
 'context': 'นายสุวัฒน์ วรรณศิริกุล (1 พฤศจิกายน พ.ศ. 2476 - 31 กรกฎาคม พ.ศ. 2555) อดีตรอง'
            'หัวหน้าพรรคพลังประชาชน อดีตประธานสมาชิกสภาผู้แทนราษฎร และประธานภาคกรุงเทพมหานคร '
            'พรรคพลังประชาชน อดีตสมาชิกสภาผู้แทนราษฎรกรุงเทพมหานครหลายสมัย ...',
 'question': 'สุวัฒน์ วรรณศิริกุล เกิดวันที่เท่าไร',
 'answers': {'text': ['1 พฤศจิกายน พ.ศ. 2476'],
             'answer_start': [24],
             'answer_end': [45]}}

This is what load_dataset returns. Earlier revisions of this card printed an example carrying three further keys — created_by, created_on and is_pay — which the loaded dataset has never contained. See "Personal and sensitive information".

Splits

	train	validation	test
questions	5,761	742	739
articles	1,529	191	192

The split is 80/10/10 at article level, so no article's questions are spread across two splits. The validation split is named validation; the file it came from was called valid.jsonl, and some older code assumes that name.

How this was built

From the original iapp-wiki-qa-dataset, @cstorm125 applied the following:

• select questions with one non-empty answer
• select questions whose answers match the textDetection fields
• select questions whose answers are 100 characters or shorter
• 80/10/10 train/validation/test split at article level

Annotators hired by iApp were asked to write questions and mark the answering span for each article. Contexts are Wikipedia authors' work.

Personal and sensitive information

The contexts are public encyclopaedia text about, in many cases, public figures. Nothing was collected about the readers of this dataset.

The annotators are a different matter, and this needs stating plainly. The records in the canonical source carry three fields that the published dataset has never exposed: created_by, a per-annotator account identifier resolving to 57 distinct people; created_on; and is_pay, which holds the payment status and amount for each annotated item. The loading script projected them away, and the parquet in this repository reproduces that projection exactly — validate.py fails if it ever stops doing so.

Earlier revisions of this card printed those three fields in its worked example, including a real account identifier and a real payment record. That has been removed. The claim in the same section that no personal information was involved was true of the six published fields and false of the example printed above it.

Verifying this card

validate.py checks the claims above against the canonical source, live:

pip install pyarrow
python validate.py

It fetches squad_format/data.zip from GitHub, applies the six-field projection itself, and compares the result against the parquet in this repository row by row. It also asserts that created_by, created_on and is_pay are absent from every file here, that every answer span is its context slice, and that no article is spread across two splits. Non-zero exit on any failure. At the current revision all 23 checks pass.

What changed, and what did not

The August 2026 revision changed the packaging and not the data. Every row, every field and every value is what the loading script produced, verified row by row against a run of that script on datasets 2.x. A score computed on this dataset before the change is comparable with one computed after it. CHANGELOG.md has the detail.

Citation

@dataset{kobkrit_viriyayudhakorn_2021_4539916,
  author       = {Kobkrit Viriyayudhakorn and Charin Polpanumas},
  title        = {iapp_wiki_qa_squad},
  month        = feb,
  year         = 2021,
  publisher    = {Zenodo},
  version      = 1,
  doi          = {10.5281/zenodo.4539916},
  url          = {https://doi.org/10.5281/zenodo.4539916}
}

Please also attribute Thai Wikipedia for the article text, as CC BY-SA requires.