CIRCLED is a multi-turn Composed Image Retrieval (CIR) dataset that addresses
two key limitations of existing datasets: lack of dialogue-history consistency
and restriction to the fashion domain. It contains 22,608 sessions across
FashionIQ, CIRR, and CIRCO datasets.
