aieng-lab/gradiend-ravel-continent

GRADIEND RAVEL Continent Data

City-to-continent cloze data adapted from RAVEL, used in aieng-lab/iend-study. A text-only neutral companion is available as aieng-lab/gradiend-ravel-neutral (config continent).

Usage

from datasets import load_dataset

ds = load_dataset("aieng-lab/gradiend-ravel-continent", split="train")

Splits: train, validation, test.

Dataset Details

Dataset Description

The prompt queries the continent associated with a city, e.g. Toronto is a city in the continent of North America. Pilibhit is a city in the continent of [MASK]. The city's true value is the factual target; a value from one of the other retained classes is the alternative. Classes: africa, asia, europe.

Dataset Structure

One config (default); 1,344 / 168 / 168 rows (train / validation / test), balanced across the three classes.

• masked: the input text with a single [MASK] slot to be predicted
• label: the factual target token(s) for the slot
• label_class: class of the factual target
• alternative: the alternative (counterfactual) target for the slot
• alternative_class: class of the alternative target
• feature_class_id: identifier of the feature class (equal to label_class)
• split: train, validation, or test
• entity, entity_id, entity_url: the queried city and its RAVEL entity
• alternative_entity, alternative_entity_id: the entity providing the alternative value
• template, cf_masked, query_attr, source, source_revision: prompt template, counterfactual text, queried attribute, and provenance

Dataset Sources

• Repository: https://github.com/aieng-lab/iend-study
• Original Data: hij/ravel (MIT), as distributed through MIB

Dataset Creation

Adapted from the RAVEL benchmark (Huang et al., 2024; hij/ravel, MIT, pinned to revision 2c45eb232b9944c235509db5320345eaba563e07) as distributed through MIB. We retain the three most frequent continent values in the RAVEL training split and sample an equal number of entities from each class. We remove city surface forms that map to multiple values of the queried attribute (e.g. Amsterdam in the Netherlands and the United States), collapse duplicate entities, and use each retained entity once with entity-disjoint splits. Sizes therefore reflect distinct, unambiguous entities rather than the repeated query/intervention pairs of the original RAVEL data.

Bias, Risks, and Limitations

Only three classes per attribute are kept, so the data is small and not representative of all cities or of the full attribute space. It inherits the coverage biases of RAVEL and its Wikipedia-derived entities.

Citation

@misc{drechsel2026gradientsinterventionsactivationsdetection,
      title={Gradients for Interventions and Activations for Detection: Targeted Feature Learning in Language Models}, 
      author={Jonathan Drechsel and Steffen Herbold},
      year={2026},
      eprint={2609.32355},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.32355}, 
}