--- dataset_info: - config_name: asian_to_black features: - name: text dtype: string - name: swapped dtype: string - name: masked dtype: string - name: source dtype: string - name: target dtype: string - name: source_id dtype: string - name: target_id dtype: string splits: - name: train num_bytes: 39396761 num_examples: 27016 - name: val num_bytes: 39396761 num_examples: 27016 - name: test num_bytes: 39396761 num_examples: 27016 download_size: 74877080 dataset_size: 118190283 - config_name: asian_to_white features: - name: text dtype: string - name: swapped dtype: string - name: masked dtype: string - name: source dtype: string - name: target dtype: string - name: source_id dtype: string - name: target_id dtype: string splits: - name: train num_bytes: 41288195 num_examples: 27016 - name: val num_bytes: 41288195 num_examples: 27016 - name: test num_bytes: 41288195 num_examples: 27016 download_size: 78565107 dataset_size: 123864585 - config_name: black_to_white features: - name: text dtype: string - name: swapped dtype: string - name: masked dtype: string - name: source dtype: string - name: target dtype: string - name: source_id dtype: string - name: target_id dtype: string splits: - name: train num_bytes: 59726209 num_examples: 36694 - name: val num_bytes: 59726209 num_examples: 36694 - name: test num_bytes: 59726209 num_examples: 36694 download_size: 114734341 dataset_size: 179178627 configs: - config_name: asian_to_black data_files: - split: train path: asian_to_black/train-* - split: val path: asian_to_black/val-* - split: test path: asian_to_black/test-* - config_name: asian_to_white data_files: - split: train path: asian_to_white/train-* - split: val path: asian_to_white/val-* - split: test path: asian_to_white/test-* - config_name: black_to_white data_files: - split: train path: black_to_white/train-* - split: val path: black_to_white/val-* - split: test path: black_to_white/test-* license: cc-by-sa-4.0 language: - en --- # GRADIEND Race Data This dataset consists of templated sentences with the masked word being sensitive to race, e.g., *African*. ``` ``` See [GENTER](https://huggingface.co/datasets/aieng-lab/genter) and [GRADIEND Religion Data](https://huggingface.co/datasets/aieng-lab/gradiend_religion_data) for similar datasets. ## Usage ```python genter = load_dataset('aieng-lab/gradiend_race_data', pair_of_races, trust_remote_code=True, split=split) ``` `split` can be either `train`, `val`, `test`, or `all`. `pair_of_races` can be `asian_to_black`, `asian_to_white`, or `black_to_white`. ## Dataset Details ### Dataset Description This dataset is a filtered version of [Wikipedia-10](https://drive.google.com/file/d/1boQTn44RnHdxWeUKQAlRgQ7xrlQ_Glwo/view?usp=sharing) 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)](https://arxiv.org/abs/2110.08527) ([bias attribute words](https://github.com/McGill-NLP/bias-bench/blob/main/data/bias_attribute_words.json)). Based on the masked term (`source`), an associated `target` is derived from a corresponding bias attribute pair, matching the casing of `source` (e.g., `White` gets to `Black` and not `black`). ### Dataset Sources - **Repository:** [github.com/aieng-lab/gradiend](https://github.com/aieng-lab/gradiend) - **Paper:** [![arXiv](https://img.shields.io/badge/arXiv-2502.01406-blue.svg)](https://arxiv.org/abs/2502.01406) - **Original Data**: [Wikipedia-10](https://drive.google.com/file/d/1boQTn44RnHdxWeUKQAlRgQ7xrlQ_Glwo) (a subset of [English Wikipedia](https://huggingface.co/datasets/wikimedia/wikipedia)) ## Dataset Structure - `text`: the original entry of Wikipedia-10 - `masked`: the masked version of `text`, i.e., with template masks for the name (`[NAME]`) and the pronoun (`[PRONOUN]`) - `swapped`: like `masked` but with inserted `target` for `[MASK]` - `source`: the word at the position of `[MASK]` in `masked` (e.g., `African`) - `source_id`: a normalized identifier for the `source` (e.g., `black`). All entries of the same `pair_of_races` have the same `source_id`. - `target`: the word inserted for `[MASK]` in `swapped` - `target_id`: a normalized identifier for the `target`. All entries of the same `pair_of_races`have the same `target_id`. ## Dataset Creation ### Curation Rationale For the training of a race bias [GRADIEND models](https://github.com/aieng-lab/gradiend), a diverse dataset is required to asses model gradients relevant to bias-sensitive information. ### Source Data The dataset is derived from [Wikipedia-10](https://drive.google.com/file/d/1boQTn44RnHdxWeUKQAlRgQ7xrlQ_Glwo) by filtering it and extracting the template structure. Whe Wikipedia-10 dump is derived from [English Wikipedia](https://huggingface.co/datasets/wikimedia/wikipedia) by [Maede et al. 2021](https://arxiv.org/pdf/2110.08527). ### 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 `pair_of_races`. ## Citation **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](https://huggingface.co/jdrechsel)