aieng-lab/gradiend_race_data

GRADIEND Race Data

This dataset consists of templated sentences with the masked word being sensitive to race, e.g., African.

See GENTER and GRADIEND Religion Data for similar datasets.

Usage

The dataset uses one subset per class. Subset names are class identifiers: white, black, asian. Each subset has columns masked, split, and one column per class (e.g. white, black, asian) giving the token for that class in that row.

from datasets import load_dataset

# Load one subset (one class view), e.g. "white"
ds = load_dataset("aieng-lab/gradiend_race_data", "white", split="train")
# ds has columns: masked, split, white, black, asian
label = ds['white']
alternative_target = ds['black'] # or 'asian'

split can be either train, val, test, or all.

Dataset Details

Dataset Description

This dataset is a filtered version of Wikipedia-10 containing only sentences that contain a race bias sensitive word of the source_id race. We used the same bias sensitive words as defined by Maede et al. (2021) (bias attribute words).

It is stored in per-class form: each subset (e.g. white) corresponds to one source class. Rows are identified by (masked, split). For each other class, the corresponding column holds the target token when that class is the counterfactual target (e.g. column black in subset white is the token used when the target class is black).

Dataset Sources

• Repository: github.com/aieng-lab/gradiend-bias
• Paper:
• Original Data: Wikipedia-10 (a subset of English Wikipedia)

Dataset Structure

• text: the original entry of Wikipedia-10
• masked: the masked version of text (i.e., contains a [MASK] at every occurrence of the subset column)
• white/asian/black: The mask target words for white/asian/black races. Note that the column equal to the subset id is the original value of the [MASK] token.

Dataset Creation

Curation Rationale

For the training of a race bias GRADIEND models, a diverse dataset is required to asses model gradients relevant to bias-sensitive information.

Source Data

The dataset is derived from Wikipedia-10 by filtering it and extracting the template structure. Whe Wikipedia-10 dump is derived from English Wikipedia by Maede et al. 2021.

Limitations

Note that the splitting is performed entirely random. Thus, the same masked text might occur in other splits (in combination with other target words). The same limitation holds across different races.

Citation

BibTeX:

@inproceedings{drechsel2026gradiend,
    title={{GRADIEND}: Feature Learning within Neural Networks Exemplified through Biases},
    author={Jonathan Drechsel and Steffen Herbold},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026},
    url={https://openreview.net/forum?id=1vBNAnAgCD}
}

Dataset Card Authors

jdrechsel