Store reviewed NCL pilot checkpoint at 375 updates

README.md

@@ -0,0 +1,21 @@
+---
+tags:
+- microscopy
+- image-segmentation
+- experimental
+---
+# NCL development pilot, 375 real-image updates
+
+Experimental NCL-staining-specific nucleolar segmentation weights, fine-tuned from an untouched cpsam_v2 checkpoint. Not adopted into the product. Not validated for NPM1 staining or cytoplasm. No synthetic images were used in this fine-tuning run.
+
+Training used 8 approved objects in 2 regions from one experiment. Development used 4 reference objects in one region from a separately cultured and treated experiment; it was never a training input. There is no completed untouched final test.
+
+Under identical local product processing, development instance F1 changed from 0.5455 to 0.8000 (TP 3 to 4, FP 4 to 2, FN 1 to 0; matching IoU at least 0.5). Extra split fragments and individual boundary/area errors remain. Raw candidate F1 is 0.3636; product postprocessing materially affects the result. Total area agreement does not establish accurate per-object quantification. Training-region geometry regressions remain versus baseline.
+
+This small development result does not establish general performance, native fluorescence measurement accuracy, or reduced human correction burden. Two-human reference agreement is not established. Preserve the base model and evaluate this candidate before any use.
+
+Only model weights and aggregate metadata are included. No research images, annotations, optimizer state, or specimen identifiers are included.
+
+Checkpoint SHA256: `bc5ea7f25ce9a9f96e9ff510cee20587e1328ac834a0f7f82aee0bc06156d906`.
+
+See `ncl/r1-step0375.json` for the fixed recipe, package versions, provenance hashes, and aggregate metrics. No additional license grant is asserted by this model card; applicable upstream terms require separate review before redistribution.

ncl/r1-step0375.json

@@ -0,0 +1,67 @@
+{
+  "schema_version": 1,
+  "task": "NCL-stained nucleolar instance segmentation",
+  "status": "experimental_development_candidate_not_adopted",
+  "completed_updates": 375,
+  "checkpoint_sha256": "bc5ea7f25ce9a9f96e9ff510cee20587e1328ac834a0f7f82aee0bc06156d906",
+  "base_checkpoint_sha256": "0f1cc3f7ecdd8a037a57c6c48d9d8921391be4cbce3fa9f13c3e3a2e1253c667",
+  "data_manifest_sha256": "0ae2728c8e85ed94d55ca0baeb54bd72c139aac01a142d7ca6ebe4e8a5a5b7fe",
+  "training_code_sha256": "4d7797c716d9fde458b29ed3b1d07e8b990cbf826560d540a62cf5c0fb94d53a",
+  "inference_code_sha256": "c4b32c7910b3e00804b48794791468dfefedec7896ec770548f34872b4fd9453",
+  "recipe": {
+    "optimizer": "AdamW",
+    "lr_max": 1e-05,
+    "weight_decay": 0.1,
+    "warmup_updates": 100,
+    "batch_size": 1,
+    "rdrop": 0,
+    "augmentation": false
+  },
+  "environment": {
+    "cellpose": "4.2.1.1",
+    "torch": "2.8.0+cpu",
+    "numpy": "2.2.6",
+    "scipy": "1.16.2"
+  },
+  "evaluation_summary": {
+    "training_experiments": 1,
+    "training_regions": 2,
+    "training_objects": 8,
+    "development_experiments": 1,
+    "development_regions": 1,
+    "development_objects": 4,
+    "final_test_regions": 0,
+    "matched_product_baseline": {
+      "TP": 3,
+      "FP": 4,
+      "FN": 1,
+      "F1": 0.5454545454545454,
+      "union_Dice": 0.8534045393858478,
+      "total_area_relative_difference": -0.2038369304556355
+    },
+    "candidate_product_output": {
+      "TP": 4,
+      "FP": 2,
+      "FN": 0,
+      "F1": 0.8,
+      "union_Dice": 0.9348565872630044,
+      "total_area_relative_difference": -0.02685851318944844
+    },
+    "candidate_raw_output": {
+      "TP": 2,
+      "FP": 5,
+      "FN": 2,
+      "F1": 0.36363636363636365,
+      "union_Dice": 0.8992209097763257,
+      "total_area_relative_difference": -0.09160671462829736
+    },
+    "limitations": [
+      "Development-only pilot; not a generalization or final validation claim.",
+      "Product output includes unchanged parent-nucleus and postprocessing stages; raw model output differs.",
+      "Extra split fragments and object-level boundary/area regressions remain.",
+      "Training images are not independent validation.",
+      "Two-human reference agreement, native fluorescence accuracy, and correction burden not measured.",
+      "Not validated for NPM1 staining or cytoplasm."
+    ]
+  }
+}

ncl/r1-step0375.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc5ea7f25ce9a9f96e9ff510cee20587e1328ac834a0f7f82aee0bc06156d906
+size 1233577155