Photos and videos contributors chose from their own camera rolls, screened for privacy, stock and AI images. Consented imagery for pretraining.

Unedited photos shot inside the Kled camera with EXIF intact, never imported from a camera roll. A camera-original class for synthetic-image detection.

Hands folding real laundry on cluttered surfaces in homes on three continents, faces out of frame. Third-person demos for deformable-object manipulation.

Read-aloud digit strings, spelled names and street addresses, one item per clip with its prompt text attached. Built for alphanumeric speech recognition.

Photographed pages of contributors' own lecture notes, problem sets and derivations, in Latin and non-Latin scripts. Real worked reasoning on paper.

Real questions typed by people who needed an answer that week, each paired with why they asked and the criteria a good answer has to meet.

Full-device screenshots of real apps at native resolution, status bar and navigation included, from Southeast Asian super-apps to US and EU apps.

Hands opening and closing doors, drawers and appliances in ordinary homes, framed tight on hand and object. Real-world articulation for home robots.

Pages contributors had already written for themselves, photographed flat from above. Hurried, multi-script handwriting in English and SEA languages.

Native speakers of Cebuano, Javanese, Sundanese and other under-resourced Southeast Asian languages, reading or speaking freely, tagged by language.

Clips from ordinary dashcams on everyday commutes, uploaded by their owners with stills alongside. Long-tail roads, from two-wheeler traffic to autobahns.

Blurred, tilted and half-cut-off photos people would normally delete, straight from their own camera rolls. Real-world failures for robustness testing.