aieng-lab/biasneutral

Dataset Card for Dataset Name

This dataset is a filtered version of BookCorpus containing only bias-neutral words (for biases gender, race, and religion).

NOTE: This dataset is derived from BookCorpus, for which we do not have publication rights. Therefore, this repository only provides indices referring to gender-neutral entries within the BookCorpus dataset on Hugging Face. By using load_dataset('aieng-lab/biasneutral', trust_remote_code=True), both the indices and the full BookCorpus dataset are downloaded locally. The indices are then used to construct the BIASNEUTRAL dataset. The initial dataset generation takes a few minutes, but subsequent loads are cached for faster access. This setup only works with datasets versions beween 3.0.2 and 3.6.0 (successfully tested with Python 3.9-3.11), as later versions deprecated trust_remote_code. If your project depends on other datasets versions, we recommend generating the dataset once in a separate Python environment, saving the resulting data to disk, and then loading the local copy within your main project.

biasneutral = load_dataset('aieng-lab/biasneutral', trust_remote_code=True, split='train')

Examples:

Index	Text
8498	no one sitting near me could tell that i was seething with rage .
8500	by now everyone knew we were an item , the thirty-five year old business mogul , and the twenty -three year old pop sensation .
8501	we 'd been able to keep our affair hidden for all of two months and that only because of my high security .
8503	i was n't too worried about it , i just do n't like my personal life splashed across the headlines , but i guess it came with the territory .
8507	i 'd sat there prepared to be bored out of my mind for the next two hours or so .
8508	i 've seen and had my fair share of models over the years , and they no longer appealed .
8512	when i finally looked up at the stage , my breath had got caught in my lungs .
8516	i pulled my phone and cancelled my dinner date and essentially ended the six-month relationship i 'd been barely having with another woman .
8518	when i see something that i want , i go after it .
8529	if i had anything to say about that , it would be a permanent thing , or until i 'd had my fill at least .

Dataset Details

• Repository: github.com/aieng-lab/gradiend
• Paper:
• Original Data: BookCorpus

Uses

This dataset is suitable for training and evaluating language models. For example, its lack of gender-related words makes it ideal for assessing language modeling capabilities in both gender-biased and gender-neutral models during masked language modeling (MLM) tasks, allowing for an evaluation independent of gender bias (same holds for race and religion bias).

Dataset Creation

We generated this dataset by filtering the BookCorpus dataset, leaving only entries matching the following criteria:

• Each entry contains at least 50 characters
• No name of aieng-lab/namextend
• No gender-specific pronoun is contained (he/she/him/her/his/hers/himself/herself)
• No gender-specific noun is contained according to the 2421 plural-extended entries of this gendered-word dataset
• No Bias Attribute Words (gender, race, religion) as defined by Maede et al. (2022).

Citation

This dataset has been originally introduced as part of GRADIEND, though it has not been used in the final version of the paper (but e.g., in version v2), as it was generalized to a more general bias neutral dataset BIASNEUTRAL.

BibTeX:

@misc{drechsel2025gradiendfeaturelearning,
      title={{GRADIEND}: Feature Learning within Neural Networks Exemplified through Biases}, 
      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