README.md

---
dataset_name: WikiCulture
language:
- en
license: mit
task_categories:
- text-generation
size_categories:
- 100K<n<1M
source_datasets:
- wikipedia

configs:
- config_name: AF
  data_files:
  - split: train
    path: AF.parquet

- config_name: AS
  data_files:
  - split: train
    path: AS.parquet

- config_name: AU
  data_files:
  - split: train
    path: AU.parquet

- config_name: CH
  data_files:
  - split: train
    path: CH.parquet

- config_name: EU
  data_files:
  - split: train
    path: EU.parquet

- config_name: LA
  data_files:
  - split: train
    path: LA.parquet

- config_name: ME
  data_files:
  - split: train
    path: ME.parquet

- config_name: NA
  data_files:
  - split: train
    path: NA.parquet
---

# Wikipedia Culture Dataset (`nDNA/WikiCulture`)

## Overview

This dataset provides coarse-grained geographic culture labels for English Wikipedia articles, mapped into eight buckets:

* **NA**: North America
* **EU**: Europe
* **AU**: Oceania (UN M49 “Oceania”)
* **AS**: Asia (excluding Western Asia and Greater China)
* **CH**: Greater China (CN, HK, MO, TW)
* **AF**: Africa
* **LA**: Latin America & Caribbean (Americas excluding Northern America)
* **ME**: Middle East (UN M49 “Western Asia”)

The labels are intended for controlled sampling and stratified analysis of Wikipedia content by broad region. They are not intended as fine-grained cultural or ethnographic ground truth.

## Data Sources

The dataset is constructed from three public components:

1. **UN M49** country/area regional classification (UN Statistics Division), used to generate a deterministic mapping from ISO3166-1 alpha-2 codes to the eight buckets.
2. **Wikipedia Cultural Diversity** dataset, used as the article-level signal source (including per-article ISO country code and Wikidata-derived fields) [https://doi.org/10.6084/m9.figshare.7039514].
3. **`wikimedia/wikipedia` (Hugging Face dataset)**, used to attach the full article text via a join.

## Labeling Methodology

### 1) Deterministic ISO2 → bucket mapping (UN M49)

An ISO3166-1 alpha-2 code is mapped to one of the eight buckets using UN M49 region/subregion fields:

* **EU** if UN M49 Region = *Europe*
* **AF** if Region = *Africa*
* **AU** if Region = *Oceania*
* **ME** if Sub-region = *Western Asia*
* **NA** if Region = *Americas* and Sub-region = *Northern America*
* **LA** if Region = *Americas* and not Northern America
* **CH** if ISO2 ∈ {CN, HK, MO, TW} (Taiwan forced to CH)
* **AS** if Region = *Asia* and not Western Asia and not CH

This mapping is deterministic and versioned.

### 2) Geo label assignment from CCC

Each article in the CCC dump provides `iso3166` (ISO3166-1 alpha-2). The primary label is:

**culture_geo = bucket(iso3166)**

Rows with missing `iso3166` or unmapped ISO2 are excluded.

### 3) High-precision consistency filtering

To increase label precision and remove cross-regional or ambiguous items, an additional consistency check is applied using CCC’s Wikidata-derived columns:

* `country_wd`
* `location_wd`

**Offline QID → ISO2 bootstrap.**
Because no online Wikidata queries are used, we construct a bootstrapped mapping from QID to ISO2 by taking the most frequent `iso3166` observed for each `qitem` in the CCC dump.

**Evidence extraction.**
For each article, QIDs are parsed from `country_wd` and `location_wd`, mapped to ISO2 using the bootstrapped table, then converted to bucket sets via the UN M49 mapping.

**Hard-strong inclusion rule.**
An article is retained as *high-precision* if:

* It has at least one non-empty evidence set from `country_wd` or `location_wd`, and
* If `location_wd` evidence is present, it must equal `{culture_geo}` exactly; otherwise,
* `country_wd` evidence must equal `{culture_geo}` exactly.

This rule is deliberately discard-heavy.

### 4) Middle East retention exception

In the CCC snapshot used, many ME-labeled rows have empty evidence after offline QID resolution (coverage limitations of the bootstrapped mapping). To avoid eliminating nearly all ME-labeled examples, a restricted exception is applied:

* If `culture_geo == "ME"` and both evidence sets are empty, the row is retained (geo-only).

## Attaching Wikipedia Text (`wikimedia/wikipedia` join)

The final dataset includes the full article text by merging the filtered labeled table with the Hugging Face dataset **`wikimedia/wikipedia`**.

* A **left inner join** is performed on the article identifier:

  * CCC / labeled table: `page_id`
  * `wikimedia/wikipedia`: `id`

Only rows that match across both sources are included in the released dataset, ensuring every labeled example has associated article text.

## Dataset Fields

Each row contains:

* `page_title` (string): Wikipedia article title (from CCC / Wikipedia metadata)
* `Qid` (string): Wikidata QID (`qitem`)
* `culture_geo` (string): one of `{NA, EU, AU, AS, CH, AF, LA, ME}`
* `text` (string): The article text (from `wikimedia/wikipedia`)

## Intended Use

This dataset is suitable for:

* constructing geographically stratified Wikipedia subsets,
* controlled pretraining mixtures or evaluation subsets by region,
* studying representation and performance disparities across broad regions.

It is not suitable for:

* fine-grained cultural identity claims,
* country-level ground truth without additional validation,
* resolving inherently transnational or multi-regional topics (which are preferentially excluded by design).

## Limitations

* Labels are coarse and operationalize “culture” as broad geography.
* The strict filter excludes many global and transnational pages.
* The offline QID→ISO2 bootstrap is a heuristic and may miss QIDs not well-represented in the CCC rows.
* The Middle East geo-only exception yields a subset with weaker evidence than the hard-strong subset, but is retained to mitigate coverage loss under offline constraints.
* The final dataset includes only articles present in both the labeled CCC-derived table and `wikimedia/wikipedia` after the `page_id` ↔ `id` join.

## Reproducibility

The pipeline is deterministic given:

* the UN M49 table snapshot used,
* the CCC dump snapshot used,
* the fixed mapping rules and filtering rule described above,
* the `wikimedia/wikipedia` snapshot/config used for the text join.

## Licensing and Attribution

This dataset is derived from Wikipedia-related resources and UN statistical classifications. Users should comply with the licensing and attribution requirements of:

* Wikipedia content (CC BY-SA and related terms),
* the CCC dataset’s licensing/terms,
* UN M49 documentation and terms where applicable.

## Citation

If you use this dataset, cite:

* the CCC dataset (Wikipedia Cultural Diversity / Cultural Context Content) and associated publication,
* UN M49 (UN Statistics Division),
* the `wikimedia/wikipedia` Hugging Face dataset,
* this dataset repository (`nDNA/WikiCulture`).