Organization for MMBU: Massive Multimodal Biomedical Understanding benchmark. Questions are full-context, open-ended visual question answering comprising the MMBU Hard subset used for evalution. Each row is one image. We encourage internal benchmarking and reporting unverified results on mmbu-public. Please reach out to us for more information on verified results on mmbu-private.
| Split | Dataset | Rows | Images |
|---|---|---|---|
| Public | mmbu-public | 3,162 | 3,162 |
| Private held-out | mmbu-private | 9,303 | 9,303 |
For the MMBU Challenge, this organization also contains the same
Task categories: classification, fine-grained classification from a detection box, fine-grained classification from a segmentation mask, object detection, and counting. Prompt format and scoring are on each dataset card.
Paper: MMBU (arXiv:2606.06696). Project page: dcunhrya.github.io/MMBU-Website.