AbstractPhil/SargeZT-coco-stuff-captioned

⚠️ WARNING: COCO Dataset Contamination Risk

The COCO dataset contains latent risks of inappropriate label associations when training adult-oriented models, particularly involving child-descriptive language.

Despite its academic origin and wide usage, COCO embeds captions such as “little girl,” “young boy,” “baby,” and “child” across a range of depictions. When used as-is in diffusion model training, this poses a serious ethical and representational hazard, as tags can be wrongly associated with adult or ambiguous content, potentially influencing generation systems toward unintended or harmful associations.