Raw inputs without ground‑truth annotations; used in self‑supervised, unsupervised, or pretraining regimes to scale model learning.