MSQA Dataset Card

MSQA (Multilingual and Multicultural SimpleQA) is a benchmark of 1,064 natively sourced questions measuring whether large language models possess genuine, locally grounded cultural knowledge — as opposed to fluency that merely looks culturally competent.

Every question was authored or curated from native, in-language sources (not translated from English), and each has a single verifiable answer.

Files

File	Description
msqa.jsonl	The benchmark, one JSON object per line (UTF-8).
msqa.csv	The same data as CSV (UTF-8 with BOM).

A single split is provided: test (1,064 items).

Fields

Field	Type	Description
id	string	Unique item id (prefixed by language, e.g. PT-01).
session_id	string	Source authoring/session id.
language	string	BCP-47-style code, e.g. pt-PT, zh-ZH, en-EN.
culture_circle	string	Cultural sphere the item targets (e.g. Portuguese, Latin American).
category	string	One of the five cultural dimensions (see below).
question	string	The question, in its native language.
answer	string	The single gold answer.
question_zh	string	Chinese translation of the question (reference aid; may be empty).
answer_zh	string	Chinese translation of the answer (reference aid; may be empty).
source_url	string	Primary source URL (may be empty).
source_url_desc	string	Short description of the source (may be empty).

Composition

Languages (11):

Language	Count		Language	Count
English (en-EN)	151		Japanese (ja-JP)	83
Chinese (zh-ZH)	150		Malay (ms-MY)	82
Thai (th-TH)	95		Indonesian (id-ID)	81
Russian (ru-RU)	92		Spanish (es-ES)	80
Korean (ko-KR)	86		Portuguese (pt-PT)	80
French (fr-FR)	84

Cultural dimensions (5):

Dimension	Count
History and Collective Memory	261
Language Expression and Communication Arts	220
Cultural Products and Symbols	208
Beliefs, Values, and Knowledge Systems	189
Social Norms and Customs	186

Note on difficulty tiers. The paper discusses three difficulty tiers (Easy/Medium/Hard). This release does not ship a per-item difficulty column; if you need tier labels, refer to the paper's appendix taxonomy.

Loading

# From the Hugging Face Hub
from datasets import load_dataset
ds = load_dataset("m-a-p/MSQA", split="test")

# From a local file (this repo)
from msqa.data import load_dataset
items = load_dataset(path="data/msqa.jsonl", language="pt-PT")

Provenance

Items were sourced from native-language references (encyclopedias, official sources, cultural documentation). source_url records the primary source where available. The clean release format is produced from the internal workbook by tools/build_dataset.py.

License

Released under CC BY 4.0. Individual items reference third-party sources via source_url; please also respect the terms of those original sources.

Citation

Paper: https://arxiv.org/abs/2607.00724

@misc{chen2026msqanativelysourcedmultilingual,
      title={MSQA: A Natively Sourced Multilingual and Multicultural SimpleQA Benchmark}, 
      author={Xianru Chen and Yukai Huang and Mingxiang Chen and Xinping Lei and Fangbing Deng and Jin Chen and Ge Zhang and Wenhao Huang and Jiaheng Liu},
      year={2026},
      eprint={2607.00724},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2607.00724}, 
}