TerraCoT

TerraCoT TerraCoT is a remote-sensing chain-of-thought VQA dataset in which every SEG token in an answer is grounded by a pixel-level ground-truth mask. Each sample pairs an optical image with a reasoning trace, and (where applicable) a second modality in image2 (Sentinel-1 SAR, or a bi-temporal companion image). Two splits: train — samples that carry an ordered labels list aligned to the SEG tokens. vqa — question/answer samples without the labels column.… See the full description on the dataset page:

Source: Hugging Face Hub (sy1998/TerraCoT). Metadata imported from the dataset’s Hub tags.