Mammo-FM Linear-probe classification artifacts
Each task directory contains the best validation-AUROC checkpoint, publication-safe SCC training log, test/validation prediction CSVs, the one-fold combined test CSV, portable JSON training configuration, TensorBoard scalars, epoch metric export, release metadata, and SHA-256 checksums.
All runs use data fraction 1.0, seed 10, fold 0, and nine epochs. Reported test AUROC uses the
dataset-specific patient/breast aggregation recorded in each task's release_metadata.json.
| Dataset | Task | Held-out test AUROC | Package |
|---|---|---|---|
| RSNA | cancer | 0.837990 | rsna/cancer_classification |
| VinDr | abnormality | 0.867237 | vindr/abnormality_classification |
| VinDr | mass | 0.779813 | vindr/mass_classification |
| VinDr | suspicious calcification | 0.974011 | vindr/suspicious_calcification_classification |
| CMMD | cancer | 0.772792 | cmmd/cancer_classification |
| CMMD | calcification | 0.868952 | cmmd/calcification_classification |
| NL-Breast | cancer | 0.826261 | nl_breast/cancer_classification |
LUMINA and CMMD mass artifacts are intentionally excluded from this release.
The released prediction tables retain public-dataset identifiers and metadata but replace local absolute filesystem prefixes with portable relative paths. SCC logs have local paths replaced with placeholders and unrelated NVIDIA process/PID tables removed. The source pickle configuration is represented as inert JSON.