N-COCO

This dataset is associated with the paper When Negation Is a Geometry Problem in Vision-Language Models.

N-COCO is a dataset designed to evaluate and improve negation understanding in Vision-Language Models (VLMs) like CLIP. The work identifies that a "negation direction" exists in the CLIP embedding space and can be manipulated via representation engineering to improve performance without fine-tuning.

• Repository: https://github.com/fawazsammani/negation-steering
• Paper: When Negation Is a Geometry Problem in Vision-Language Models

Dataset Description

The dataset includes the following files used for training and evaluation:

• train_data.json: Contains positive-negative sentence pairs used to train layerwise steering directions.
• simpleneg.json: An evaluation benchmark for negation understanding.

Citation

If you find this work or dataset useful, please consider citing:

@misc{sammani2026negationgeometry,
  title={When Negation Is a Geometry Problem in Vision-Language Models},
  author={Fawaz Sammani and Tzoulio Chamiti and Paul Gavrikov and Nikos Deligiannis},
  year={2026},
  eprint={2603.20554},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}