NAMEXACT

This dataset contains names that are exclusively associated with a single gender and that have no ambiguous meanings, therefore being exact with respect to both gender and meaning.

The data is split into train, validation, and test set. You can load the entire dataset using:

from datasets import load_dataset
load_dataset('aieng-lab/genter', split='all')

Dataset Details

Dataset Description

The goal of this dataset to consist only of words that are clearly names of unabiguous gender. For instance, the following names are excluded:

• Skyler (ambiguous gender)
• Christian (believer in Christianity)
• Drew (simple past of the verb to draw)
• Florence (an Italian city)
• Henry (the SI unit of inductance)
• Mercedes (a car brand)

Due to the exclusion of such names, this dataset can be used for tasks where only names (with high certainty) are required.

A larger name dataset is NAMEXTEND.

Dataset Sources [optional]

• Repository: github.com/aieng-lab/gradiend-bias
• Paper:
• Original Dataset: Gender by Name

Dataset Structure

This dataset comes in a version containing all names (split), and three splits: train (85%), validation(5%), test (10%)

• name: the name
• gender: the gender of the name (M for male and F for female)
• count: the count value of this name (raw value from the original dataset)
• probability: the probability of this name (raw value from original dataset; not normalized to this dataset!)
• split: the split of the name (constant for HuggingFace splits train/ validation/ test; but contains the respective HuggingFace splits for all)

Dataset Creation

Source Data

The data is created by filtering Gender by Name.

Data Collection and Processing

First, all names of the raw dataset with counts less than 20000 are filtered out, resulting in a selection of the most common 1697 names. Next, we removed names with ambiguous gender, such as Skyler, Sidney, and Billie, which were identified by having counts for both genders in the filtered dataset, removing 67 additional names.

To further refine our selection of the remaining 1,630 names, we manually checked each remaining name for ambiguous meanings, such as Christian (believer in Christianity), and Drew (simple past of the verb to draw). This exclusion process was performed without considering casing to ensure applicability to non-cased models. The filtering resulted in the exclusion of 232 names, leaving us with a total of 1398 names in this dataset NAMEXACT.

Bias, Risks, and Limitations

The original dataset provides counts of names (with their gender) for male and female babies from open-source government authorities in the US (1880-2019), UK (2011-2018), Canada (2011-2018), and Australia (1944-2019) in these periods.

Citation

BibTeX:

@misc{drechsel2025gradiendmonosemanticfeaturelearning,
      title={{GRADIEND}: Monosemantic Feature Learning within Neural Networks Applied to Gender Debiasing of Transformer Models}, 
      author={Jonathan Drechsel and Steffen Herbold},
      year={2025},
      eprint={2502.01406},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2502.01406}, 
}

Dataset Card Authors

jdrechsel