"Would you please look at the data?": giffmana flags bad annotations in FineVision's coco_color

giffmana · x · 2026-11-21

Google DeepMind researcher Matthias Minderer (giffmana) showcases examples from cococolor, part of the FineVision dataset, questioning its annotation quality: a beach pizza described as "black, brown, tan tones"; giraffes captioned with imaginary zebras in "lightblue, gray, black" tones; and a bedroom scene with wildly wrong color descriptions. The post underscores how noisy auto-generated QA pairs in multimodal training data can introduce systematic errors.

Original post →

More from Multimodal

Multimodal channel →