aieng-lab/gradiend_wiki_gender

WikiGender (GRADIEND – Wikipedia Gender)

WikiGender is a masked evaluation dataset derived from the English Wikipedia.
It contains naturally occurring sentences with exactly one gendered target word, which is replaced by [MASK].

The dataset is designed to evaluate whether GRADIEND-style encoders and masked language models generalize gender-related representations to:

• unseen lexical items
• a different domain than the training source

---

Target Words

The following gendered word pairs are included:

• she / he
• woman / man
• girl / boy
• mother / father
• daughter / son

Each example contains exactly one of these tokens.

For each target token, up to 1,000 masked texts are retained.

---

Dataset Construction

Source

• Dataset: wikimedia/wikipedia
• Configuration: 20231101.en
• Language: English
• License: CC-BY-SA (Wikipedia)
• Repository: https://github.com/aieng-lab/gradiend-bias
• Paper: https://arxiv.org/abs/2502.01406

Processing Pipeline

From the full Wikipedia dump:

1. Text is Unicode-normalized (NFC).
2. Only examples containing exactly one target token are kept.
3. The token is replaced with [MASK] (first occurrence only).
4. Examples where [MASK] appears within the first 10 words are removed.
5. The dataset is balanced to ≤ 1,000 examples per token.

Not all intermediate filtered examples are published.
Only the final balanced subset is uploaded to Hugging Face.

Repository: https://github.com/aieng-lab/gradiend-bias

---

Dataset Structure

Each entry contains:

Field	Description
url	Wikipedia URL
title	Wikipedia article title
text	Original Wikipedia text
masked_text	Text with the gendered token replaced by [MASK]
token	The original gendered word (e.g., 'man')
label	Binary gender label ('M' for male, 'F' for female)

Subsets

• pronoun: he/she
• adult_noun: man/woman
• child_noun: boy/girl
• adult_role: father/mother
• child_role: son/daughter

---

Example

Original She was born in London and later became a professor of mathematics. Masked [MASK] was born in London and later became a professor of mathematics. Metadata token: she; group: pronoun; label: M

---

Intended Use

WikiGender is intended for:

• Evaluating masked language models (MLM)
• Measuring gender-related gradient signals
• Testing generalization to unseen lexical items
• Comparing semantic categories (pronouns vs. kinship terms vs. person nouns)
• Bias analysis across domains

The dataset is particularly suited for single-token masked prediction tasks.

---

Usage

from datasets import load_dataset

dataset = load_dataset("aieng-lab/gradiend_wiki_gender", "pronouns", split="train")

---

Relation to Other GRADIEND Datasets

WikiGender complements:

• GENTER (BookCorpus-based name–pronoun templates)
• GRADIEND Race Data
• GRADIEND Religion Data

---

Limitations

• Focuses on binary gender only
• Gender is inferred purely from lexical items
• Wikipedia reflects historical and societal biases
• Context may still contain implicit gender signals

---

Citation

If you use this dataset, please cite:

@inproceedings{drechsel2026gradiend,
  title     = {{GRADIEND}: Feature Learning within Neural Networks Exemplified through Biases},
  author    = {Drechsel, Jonathan and Herbold, Steffen},
  booktitle = {Proceedings of the International Conference on Learning Representations},
  year      = {2026},
  url       = {https://arxiv.org/abs/2502.01406}
}