M4 Bias Eval FairFace

M4 Bias Eval FairFace is a dataset for evaluating social bias in vision-language models, which generate text from images. It was created by Hugging Face's M4 team to assess bias in IDEFICS, its open multimodal model, and was released alongside the model's documentation. The dataset shows whether a model's descriptions of people vary in stereotypical ways depending on their perceived gender, ethnicity or age. It is released under a CC BY 4.0 licence.

The dataset contains around 11,000 face images drawn from FairFace, an existing dataset designed to be balanced across demographic groups. Each image is labelled with a perceived gender, one of seven ethnicity categories and one of nine age ranges. Each image was shown to two versions of IDEFICS, with 9 billion and 80 billion parameters. Each model was asked to describe the person and then complete three open-ended tasks: write a résumé for them, write a dating profile in their voice, and write a news article about their arrest.

The dataset records all six generated texts for each image alongside its demographic labels. Researchers can then compare the outputs across groups. For example, they can examine which professions and qualifications appear in résumés, or what crimes are described in arrest articles, for each gender, ethnicity or age range. Differences point to stereotypes the models have learned. The same method can be applied to other vision-language models.

Auto-discovered on 2026-07-15 by OECD Catalogue Automation