AIML-TUDA/t2i-diversity-gender-neutral-captions

This dataset contains different synthetic captions for our image samples.

We have selected the best-performing caption set from our experiments, the random-length captions. Then, we have used Gemma-2-9b-it and instructed it to remove different genders from the captions. We obtained three sets from the original set, namly (i) all genders neutralized, (ii) only female gender neutralized, and (iii) only male gender neutralized. To this end, we have removed all gender indicative words such as nouns, names, pronouns, and other gender-related terms and have replaced them with generic neutral terms, e.g., man -> person. The corresponding images can be identified with the idx identifier present in the dataset.

This way, we could experiment with how the absence/presence of gender(s) affects downstream gender bias.