| license: other | |
| task_categories: | |
| - visual-question-answering | |
| - image-segmentation | |
| tags: | |
| - breast-ultrasound | |
| - biomedclip | |
| - gmpo | |
| - vqa | |
| # BUS-COT Training Bundle | |
| This public dataset repository packages the training inputs and checkpoints used | |
| for the BUS-COT breast-ultrasound experiments. | |
| ## Contents | |
| - `data/training_csvs/`: GMPO and positive-only CLIP training CSV files. | |
| - `data/image_assets/`: positive images, GMPO negative images, and lesion masks. | |
| - `data/vqa/full_train_and_test_data/`: BUS-COT VQA train/test JSON files. | |
| - `checkpoints/upstream/`: selected GMPO `oa25_like_012` epoch-3 checkpoint. | |
| - `manifests/`: GMPO-selection and evaluation summaries. | |
| ## Selected GMPO | |
| The only uploaded checkpoint is GMPO `oa25_like_012`, epoch 3, with global | |
| four-way weights `[0.5, 0.5, 1.0, 0.25]`. | |
| ## Important Notes | |
| - The UDIAT segmentation result was used to choose among GMPO candidates; it is | |
| not an independent blind-final-test estimate. | |
| - Users are responsible for verifying the source data licence, consent status, | |
| institutional approval, and suitability for their intended use. | |