Add LUMINA cancer classification checkpoints

Add linear-probe and full-fine-tuning inference checkpoints with integrity and input metadata.

fullfinetune_mammo_FM_image_encoder/lumina/cancer_classification/README.md

@@ -0,0 +1,9 @@
+# LUMINA cancer classification — Full fine-tuning
+
+Mammo-FM EfficientNet-B5 image encoder with a binary cancer-classification head.
+
+The checkpoint was selected by validation AUROC from a completed nine-epoch run using the full training split (seed 10, fold 0). Input geometry is 1520 pixels high × 912 pixels wide, with three image channels. The model returns one logit per image; apply sigmoid for its cancer score.
+
+Checkpoint: [`mammo_fm_lumina_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_lumina_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth). Integrity and input metadata are in [`release_metadata.json`](release_metadata.json).
+
+The checkpoint contains model tensors and simple inference metadata only. It can be loaded with `torch.load(..., weights_only=True)` or the Mammo-FM strict classifier backend. For research use; not a standalone clinical device.

fullfinetune_mammo_FM_image_encoder/lumina/cancer_classification/checkpoint/mammo_fm_lumina_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d8ace59af336932ac3887f5028ae5923a837764128dba70ebb5d29e80ef5e06
+size 114438790

fullfinetune_mammo_FM_image_encoder/lumina/cancer_classification/release_metadata.json

@@ -0,0 +1,30 @@
+{
+  "release_schema_version": 1,
+  "dataset": "LUMINA",
+  "task": "cancer",
+  "method": "full_finetune",
+  "model": "Mammo-FM EfficientNet-B5 image encoder with binary classification head",
+  "checkpoint_file": "checkpoint/mammo_fm_lumina_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth",
+  "checkpoint_sha256": "7d8ace59af336932ac3887f5028ae5923a837764128dba70ebb5d29e80ef5e06",
+  "checkpoint_bytes": 114438790,
+  "checkpoint_selection": "best validation AUROC",
+  "selected_epoch": 4,
+  "training_epochs": 9,
+  "seed": 10,
+  "fold": 0,
+  "data_fraction": 1.0,
+  "image_size_hw": [
+    1520,
+    912
+  ],
+  "input_channels": 3,
+  "feature_dimension": 2048,
+  "output_classes": 1,
+  "output_activation": "sigmoid",
+  "normalization_mean": 0.3089279,
+  "normalization_std": 0.25053555408335154,
+  "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
+  "model_tensor_count": 854,
+  "strict_mammofm_load": "PASS",
+  "source_model_tensor_equality": "PASS"
+}

linear_probe_mammo_FM_image_encoder/lumina/cancer_classification/README.md

@@ -0,0 +1,9 @@
+# LUMINA cancer classification — Linear probe
+
+Mammo-FM EfficientNet-B5 image encoder with a binary cancer-classification head.
+
+The checkpoint was selected by validation AUROC from a completed nine-epoch run using the full training split (seed 10, fold 0). Input geometry is 1520 pixels high × 912 pixels wide, with three image channels. The model returns one logit per image; apply sigmoid for its cancer score.
+
+Checkpoint: [`mammo_fm_lumina_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_lumina_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth). Integrity and input metadata are in [`release_metadata.json`](release_metadata.json).
+
+The checkpoint contains model tensors and simple inference metadata only. It can be loaded with `torch.load(..., weights_only=True)` or the Mammo-FM strict classifier backend. For research use; not a standalone clinical device.

linear_probe_mammo_FM_image_encoder/lumina/cancer_classification/checkpoint/mammo_fm_lumina_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:416b7eff952bff6101480e1a074188a243a74009cb4d75b13eb5f1f551821fa6
+size 114434220

linear_probe_mammo_FM_image_encoder/lumina/cancer_classification/release_metadata.json

@@ -0,0 +1,30 @@
+{
+  "release_schema_version": 1,
+  "dataset": "LUMINA",
+  "task": "cancer",
+  "method": "linear_probe",
+  "model": "Mammo-FM EfficientNet-B5 image encoder with binary classification head",
+  "checkpoint_file": "checkpoint/mammo_fm_lumina_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth",
+  "checkpoint_sha256": "416b7eff952bff6101480e1a074188a243a74009cb4d75b13eb5f1f551821fa6",
+  "checkpoint_bytes": 114434220,
+  "checkpoint_selection": "best validation AUROC",
+  "selected_epoch": 8,
+  "training_epochs": 9,
+  "seed": 10,
+  "fold": 0,
+  "data_fraction": 1.0,
+  "image_size_hw": [
+    1520,
+    912
+  ],
+  "input_channels": 3,
+  "feature_dimension": 2048,
+  "output_classes": 1,
+  "output_activation": "sigmoid",
+  "normalization_mean": 0.3089279,
+  "normalization_std": 0.25053555408335154,
+  "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
+  "model_tensor_count": 854,
+  "strict_mammofm_load": "PASS",
+  "source_model_tensor_equality": "PASS"
+}