Dataset curation and review rubric
Three reported curation dimensions
The manuscript describes three curation dimensions. A record is appropriate for retention only when all three dimensions are satisfactory.
| Dimension | Review question | Pass condition |
|---|---|---|
| Question validity | Is the instruction clinically meaningful, clear, and grounded in features visible in the input image? | The requested task is unambiguous and does not require information absent from the supplied input. |
| Answer accuracy | Does the answer agree with the reference image and the available record metadata? | The answer uses appropriate medical terminology and describes the reference output without unsupported claims. |
| Multimodal relevance | Is the image necessary to resolve the instruction? | The task cannot be answered adequately from wording alone; image content determines the answer or transformation. |
Review-attachment selection checklist
This checklist was applied to select the review-package examples. It documents package selection and does not substitute for a retrospective human-expert calibration study.
- Input image(s) and reference image(s), when applicable, open correctly and are visually legible at native resolution.
- The task, modality, instruction, answer, and available metadata are mutually consistent enough for reviewer inspection.
- The selected records cover all 15 named tasks and the six canonical reporting modalities: CT, MRI, ultrasound, X-ray, pathology, and clinical photography.
- Clinical-photograph examples are non-identifying cropped skin images; no names, faces, or account identifiers are supplied in this package.
- Counterfactual generation/editing examples are explicitly labeled as synthetic and non-diagnostic.
- No source-data path, author identifier, API credential, or repository-account identifier is retained in the distributed JSON records.