Replace incorrect risk weights with five BU S1 KD plus S2 benchmark checkpoints

risk_checkpoints/{Mammo-FM_risk_stage1_distillation_fold1.pth-best.tar → BU_S1_KD_S2/fold1/Mammo-FM_BU_S1_KD_S2_fold1.pth-best.tar}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bd16aa6dba9f1812d52947a325df92cd8aace440ddec2fee21cbd8357fcdf41a
+oid sha256:f57730d674716c5870e5f0678c82c2a21cdff37eea518059f09809032830653e
 size 184831762

risk_checkpoints/{Mammo-FM_risk_stage2_finetuning_fold1.pth-best.tar → BU_S1_KD_S2/fold2/Mammo-FM_BU_S1_KD_S2_fold2.pth-best.tar}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:397098ff49c914dba52a56c9c3ec530fe4d1637170a0a536d5e84c0dc95bc6cf
+oid sha256:cac0884effa796de7fc32ac98b630301755d1d72a12a52addab24853a2c618a9
 size 184831762

risk_checkpoints/BU_S1_KD_S2/fold3/Mammo-FM_BU_S1_KD_S2_fold3.pth-best.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6b1cf5845c708a530ef8f5f98ecf2db87b54217f74b430e9b068ec99a40d1d6
+size 184831762

risk_checkpoints/BU_S1_KD_S2/fold4/Mammo-FM_BU_S1_KD_S2_fold4.pth-best.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1bf8769f3d6ee86253e36a473047396fb2e9d34470c440b8c5c4421dcbb9742
+size 184831762

risk_checkpoints/BU_S1_KD_S2/fold5/Mammo-FM_BU_S1_KD_S2_fold5.pth-best.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0df00131c5a91ccd2e9d464bdbfb561f9604d589e38e61fa2903de5c50b8490f
+size 184831762

risk_checkpoints/BU_S1_KD_S2/manifest.json

@@ -0,0 +1,46 @@
+{
+  "schema_version": 1,
+  "model": "Mammo-FM BU S1 KD + S2",
+  "training_variant": "from_stage1",
+  "description": "Five final Stage-2 risk models initialized from Stage-1 knowledge distillation. Each model is evaluated only on its matching held-out BU patient fold.",
+  "outputs": "Five raw independent sigmoid probabilities; uncalibrated.",
+  "metric_aggregation": "Unweighted mean across five held-out folds; AUPRC uses average precision and year-specific eligibility masks.",
+  "files": [
+    {
+      "fold": 1,
+      "path": "risk_checkpoints/BU_S1_KD_S2/fold1/Mammo-FM_BU_S1_KD_S2_fold1.pth-best.tar",
+      "sha256": "f57730d674716c5870e5f0678c82c2a21cdff37eea518059f09809032830653e",
+      "bytes": 184831762,
+      "stored_epoch": 2
+    },
+    {
+      "fold": 2,
+      "path": "risk_checkpoints/BU_S1_KD_S2/fold2/Mammo-FM_BU_S1_KD_S2_fold2.pth-best.tar",
+      "sha256": "cac0884effa796de7fc32ac98b630301755d1d72a12a52addab24853a2c618a9",
+      "bytes": 184831762,
+      "stored_epoch": 4
+    },
+    {
+      "fold": 3,
+      "path": "risk_checkpoints/BU_S1_KD_S2/fold3/Mammo-FM_BU_S1_KD_S2_fold3.pth-best.tar",
+      "sha256": "f6b1cf5845c708a530ef8f5f98ecf2db87b54217f74b430e9b068ec99a40d1d6",
+      "bytes": 184831762,
+      "stored_epoch": 6
+    },
+    {
+      "fold": 4,
+      "path": "risk_checkpoints/BU_S1_KD_S2/fold4/Mammo-FM_BU_S1_KD_S2_fold4.pth-best.tar",
+      "sha256": "b1bf8769f3d6ee86253e36a473047396fb2e9d34470c440b8c5c4421dcbb9742",
+      "bytes": 184831762,
+      "stored_epoch": 6
+    },
+    {
+      "fold": 5,
+      "path": "risk_checkpoints/BU_S1_KD_S2/fold5/Mammo-FM_BU_S1_KD_S2_fold5.pth-best.tar",
+      "sha256": "0df00131c5a91ccd2e9d464bdbfb561f9604d589e38e61fa2903de5c50b8490f",
+      "bytes": 184831762,
+      "stored_epoch": 5
+    }
+  ],
+  "validation_status": "Historical benchmark checkpoints identified and file hashes verified; fresh port inference validation is separate and pending."
+}

risk_checkpoints/README.md

@@ -1,47 +1,96 @@
-# Mammo-FM risk checkpoints
-
-The historical risk artifacts are released at
-[`batmanLab/Mammo-FM/risk_checkpoints`](https://huggingface.co/batmanLab/Mammo-FM/tree/main/risk_checkpoints).
-
-| File | Original training stage | Payload epoch | Bytes | SHA-256 |
-| --- | --- | ---: | ---: | --- |
-| `Mammo-FM_risk_stage1_distillation_fold1.pth-best.tar` | MIRAI probability distillation | 17 | 184,831,762 | `bd16aa6dba9f1812d52947a325df92cd8aace440ddec2fee21cbd8357fcdf41a` |
-| `Mammo-FM_risk_stage2_finetuning_fold1.pth-best.tar` | supervised five-year risk training | 8 | 184,831,762 | `397098ff49c914dba52a56c9c3ec530fe4d1637170a0a536d5e84c0dc95bc6cf` |
-
-These files are uploaded byte-for-byte under descriptive release names. Their
-payloads contain `model`, `optimizer`, `scheduler`, and `epoch`. They do not
-contain explicit `stage`, `fold`, validation loss, class weights, teacher
-provenance, or base-model configuration. “fold1” is retained from the original
-filename rather than inferred from checkpoint metadata.
-
-The Stage 2 training run has separate local completion evidence at epoch 20
-(SHA-256 `5238eff89cae22106f4c35f23434ba448190a68232a1214ceaedce19c5268cab`);
-that latest-state file is not part of this best-checkpoint release. No
-corresponding Stage 1 completion artifact was located, so Stage 1 is described
-only as the historical best-validation checkpoint at embedded epoch 17.
-
-The historical model state contains 1,763 entries because its wrapper
-registered the same EfficientNet image encoder at both `image_encoder.*` and
-`aggregator.image_encoder.*`. The two namespaces are duplicates within each
-file. `mayo_mammofm_risk.model.compatible_model_state` verifies tensor equality
-and removes only the duplicate namespace, leaving 911 keys for strict loading.
-
-The original Mayo launchers explicitly identify these exact filenames as the
-Stage 1 and Stage 2 Mayo models. The payloads themselves do not contain enough
-metadata to independently re-establish that lineage, so the original path,
-filename, payload inventory, and hashes are retained as provenance. Between
-the two checkpoints, all image-encoder parameter tensors remain identical;
-only BatchNorm running means, variances, and counters differ. This matches the
-historical wrapper leaving frozen BatchNorm modules in training mode. The
-cleaned runner corrects that behavior and therefore may not reproduce the
-historical numbers bit-for-bit.
-
-The historical optimizer state is not resumed by the cleaned runner. Only
-model weights are loaded. A compatible Mammo-FM base checkpoint is also
-required to construct the image-encoder configuration before loading either
-file. Strict loading was verified with the public
-[`Mammo-FM_BatmanlabTrained_CLIP.tar`](https://huggingface.co/batmanLab/Mammo-FM/blob/main/Mammo-FM_BatmanlabTrained_CLIP.tar)
-checkpoint: the normalized Stage 2 state matched all 911 model keys.
+# Mammo-FM risk checkpoints: BU S1 KD + S2
+
+This release contains five BU fold-specific **final Stage 2 risk models
+initialized through Stage 1 MIRAI knowledge distillation**. It is the
+historical `all_years/aug_n/from_stage1` model family, not Stage 1 alone and
+not the S2-only comparison variant. These checkpoints replace the previously
+released Mayo-labelled Stage 1/Stage 2 pair, which was not the intended BU
+benchmark release.
+
+## Download and provenance
+
+All five files contain 184,831,762 bytes. Stored epoch values are reproduced
+without changing their indexing convention. The
+[release manifest](https://huggingface.co/batmanLab/Mammo-FM/blob/main/risk_checkpoints/BU_S1_KD_S2/manifest.json)
+contains the same checkpoint identities and hashes.
+
+| BU fold | Final S1 KD + S2 checkpoint | Stored epoch | SHA-256 |
+| --- | --- | ---: | --- |
+| 1 | [Download fold 1](https://huggingface.co/batmanLab/Mammo-FM/resolve/main/risk_checkpoints/BU_S1_KD_S2/fold1/Mammo-FM_BU_S1_KD_S2_fold1.pth-best.tar) | 2 | `f57730d674716c5870e5f0678c82c2a21cdff37eea518059f09809032830653e` |
+| 2 | [Download fold 2](https://huggingface.co/batmanLab/Mammo-FM/resolve/main/risk_checkpoints/BU_S1_KD_S2/fold2/Mammo-FM_BU_S1_KD_S2_fold2.pth-best.tar) | 4 | `cac0884effa796de7fc32ac98b630301755d1d72a12a52addab24853a2c618a9` |
+| 3 | [Download fold 3](https://huggingface.co/batmanLab/Mammo-FM/resolve/main/risk_checkpoints/BU_S1_KD_S2/fold3/Mammo-FM_BU_S1_KD_S2_fold3.pth-best.tar) | 6 | `f6b1cf5845c708a530ef8f5f98ecf2db87b54217f74b430e9b068ec99a40d1d6` |
+| 4 | [Download fold 4](https://huggingface.co/batmanLab/Mammo-FM/resolve/main/risk_checkpoints/BU_S1_KD_S2/fold4/Mammo-FM_BU_S1_KD_S2_fold4.pth-best.tar) | 6 | `b1bf8769f3d6ee86253e36a473047396fb2e9d34470c440b8c5c4421dcbb9742` |
+| 5 | [Download fold 5](https://huggingface.co/batmanLab/Mammo-FM/resolve/main/risk_checkpoints/BU_S1_KD_S2/fold5/Mammo-FM_BU_S1_KD_S2_fold5.pth-best.tar) | 5 | `0df00131c5a91ccd2e9d464bdbfb561f9604d589e38e61fa2903de5c50b8490f` |
+
+```bash
+hf download batmanLab/Mammo-FM \
+  --include 'risk_checkpoints/BU_S1_KD_S2/*' \
+  --local-dir /path/to/checkpoints
+```
+
+The weights are released byte-for-byte from the historical best-validation
+checkpoints. Their payloads contain `model`, `optimizer`, `scheduler`, and
+`epoch`; fold and training-stage lineage come from the corresponding original
+experiment artifacts. No images, reports, patient-level tables, predictions,
+feature caches, or API credentials are included. Store downloaded weights
+and all experiment outputs outside the source repository.
+
+## BU held-out evaluation
+
+Checkpoint `fold{k}` must be evaluated only on source rows with
+`split_fold_{k}=test_fold_{k}`. The five patient-disjoint test folds cover
+13,694 examinations from 9,900 patients. They are not an ensemble; do not
+apply one fold's model to the entire cohort or average the five models'
+predictions for this benchmark.
+
+Preserve the historical BU PNG preprocessing and the four-view order
+`LMLO -> LCC -> RMLO -> RCC`, including its view/side metadata. A generic
+Mayo-style DICOM loader is a different input pipeline and is not a drop-in
+replacement for this benchmark.
+
+The model returns raw, uncalibrated independent sigmoid probabilities for
+years 1 through 5. AUROC and AUPRC (average precision) are evaluated within
+each fold using horizon-specific follow-up masks and reported as unweighted
+five-fold means. C-index is reported as the mean and between-fold standard
+deviation. Do not replace these with pooled metrics, calibrated risks, or
+monotonic/cumulative post-processing.
+
+| Horizon | Eligible examinations | Positive examinations |
+| --- | ---: | ---: |
+| 1 year | 13,694 | 283 |
+| 2 years | 12,724 | 495 |
+| 3 years | 11,878 | 700 |
+| 4 years | 11,113 | 888 |
+| 5 years | 10,369 | 1,044 |
+
+## Model loading
+
+The historical model state registers the same encoder under
+`image_encoder.*` and `aggregator.image_encoder.*`. The Mammo-FM compatibility
+loader verifies exact equality of these duplicate tensors before removing
+only the duplicate namespace. The complete remaining state must then be
+loaded strictly.
+
+Download the compatible base checkpoint for architecture construction:
+
+```bash
+hf download batmanLab/Mammo-FM Mammo-FM_BatmanlabTrained_CLIP.tar \
+  --local-dir /path/to/checkpoints
+```
+
+The base checkpoint supplies the encoder architecture configuration. Loading
+the full risk checkpoint replaces the encoder and all risk-model tensors;
+the final predictions do not retain the base encoder's initialization.
+Optimizer and scheduler states are not used for inference. Load serialized
+training checkpoints only from a trusted source.
+
+## Validation status and use restrictions
+
+Historical checkpoint identity and ported-code numerical parity are separate
+checks. Availability of these weights does not establish that the port has
+reproduced the historical BU held-out predictions or metrics. The fold-matched
+comparison must complete before such a claim is made.
 
 The repository's model-weight license applies. The weights are for
-non-commercial academic research and must not be used for clinical decisions.
+non-commercial academic research only and must not be used for clinical
+decisions.