CounterFact-GNM

Dataset created for for the paper Tell Me What To Learn: Generalizing Neural Memory to be Controllable in Natural Language" (Bennett M, Zollo T, Zemel R, 2026).

It is a categorized and document-augmented version of the CounterFact dataset.

Dataset Description

This dataset extends CounterFact by:

• Category labels for each fact
• Synthetic documents containing the 2-8 different facts
• Train/validation/test splits based on held-out categories to test for generalization

Citation

If you use this dataset, please cite both our work and the original CounterFact paper.

@misc{bennett2026gnm,
  title={Tell Me What To Learn: Generalizing Neural Memory to be
Controllable in Natural Language},
  author={Bennett M, Zollo T, Zemel R},
  booktitle={ArXiv},
  year={2026}
}

@inproceedings{meng2022locating,
  title={Locating and Editing Factual Associations in GPT},
  author={Meng, Kevin and others},
  booktitle={NeurIPS},
  year={2022}
}```