Century: A Dataset of Sensitive Historical Images

Abstract

How do we measure the way multi-modal generative models, like GPT-4 and Gemini, describe images of historical events and figures, whose legacies may be nuanced, multifaceted, or contested? As a first step to addressing this challenge, we introduce Century – a novel dataset of sensitive historical images. This dataset consists of 1,500 images from recent history, created through a novel automated method combining knowledge graphs and language models, while being rooted in the practices of museums and digital archives. We demonstrate through automated and human evaluation that this method produces a set of images that depict events and figures that are diverse across topics and represents all regions of the world, with implications for the development of evaluations for historical contextualisation and socio-cultural understanding.