FINER-DOCCI

FINER-DOCCI contains the text annotations for the FINER-DOCCI benchmark.

This repository does not include DOCCI images. The image column stores the original DOCCI image filename, e.g. test_00000.jpg. Users should download the DOCCI images separately and map each row to the corresponding image by filename.

Splits

• multi_obj
• multi_attr
• multi_rel
• wh

Notes

FINER-DOCCI is designed as a large-scale, open-set stress-test benchmark derived from dense human captions. Compared with FINER-CompreCap, it prioritizes scale and open-set coverage rather than fully human-verified per-sample annotation fidelity.