Serve single-photo head with vetted taxonomy corrections

Dockerfile

@@ -28,9 +28,9 @@ COPY . .
 # Pinned single-photo head, legacy sex detector and checklist. The reviewed Wings v6
 # detector is a Space asset and must not be overwritten.
 ARG WEIGHTS_REPO=fr4nzzch/butterfly-id-classifier
-ARG WEIGHTS_REVISION=047298dd04d0b3c3cf3f6a5b68fc0f07046b7d8f
+ARG WEIGHTS_REVISION=173f2b75ea834274120b9217055f1e763f7945d7
 RUN mkdir -p /app/assets && python -c "import shutil; from huggingface_hub import hf_hub_download; [shutil.copy(hf_hub_download('${WEIGHTS_REPO}', f, revision='${WEIGHTS_REVISION}'), '/app/assets/'+f) for f in ['head_hier.pt','wing_seg.pt','region_checklist.json']]"
-RUN python -c "import hashlib,pathlib; p=pathlib.Path('/app/assets'); expected={'head_hier.pt':'bc3c6275a69b6391b2520d53b267f011a9ab1405991dd8e2ab4be02b8b44c1ef','wing_seg_v6.pt':'38a0dab20ee607e0a43a72488420ed50e329f07c71c95775ef647b19e0f275de','field_wingtip_override_v1.json':'fb65b55b1f4ae26ce22383f2fcf216d41e104ebc47edc4db6aa82094718a6cb1'}; [(hashlib.sha256((p/n).read_bytes()).hexdigest()==sha or (_ for _ in ()).throw(ValueError('asset hash mismatch '+n))) for n,sha in expected.items()]"
+RUN python -c "import hashlib,pathlib; p=pathlib.Path('/app/assets'); expected={'head_hier.pt':'fe6f00821dc06cace5238d71c2561cc6eae37bcf1616129f65da8ccba5583eea','wing_seg_v6.pt':'38a0dab20ee607e0a43a72488420ed50e329f07c71c95775ef647b19e0f275de','field_wingtip_override_v1.json':'fb65b55b1f4ae26ce22383f2fcf216d41e104ebc47edc4db6aa82094718a6cb1'}; [(hashlib.sha256((p/n).read_bytes()).hexdigest()==sha or (_ for _ in ()).throw(ValueError('asset hash mismatch '+n))) for n,sha in expected.items()]"
 
 # Bake the 4 GB BioCLIP 2.5-H backbone into the in-image HF cache (HF_HOME) at BUILD
 # time. This is the big cold-start win: a freshly woken container loads the backbone

README.md

@@ -11,9 +11,10 @@ license: mit
 
 # Butterfly ID API
 
-This Space serves the published BioCLIP-2.5-H image encoder and the 7,841-leaf
-single-photo taxonomy head, retrained on 27 September 2026 with 558 corrected
-training labels and added GBIF photographs (up to 1,000 observations per species). `/metadata` reports the active pipeline revision and SHA-256
+This Space serves the published BioCLIP-2.5-H image encoder and the 7,482-leaf
+single-photo taxonomy head, retrained on 27 September 2026 with added GBIF
+photographs (up to 1,000 observations per species) and 757 vetted taxonomy corrections.
+Leaf names follow the corrected taxonomy; the old-to-new map is `taxon_name_map.json` in the model repo. `/metadata` reports the active pipeline revision and SHA-256
 identifiers for the head and Wings v6 detector. Check those identifiers before using a new response contract.
 
 For taxonomy, the API runs the released Wings v6 detector on raw image pixels,
@@ -36,7 +37,7 @@ legacy detector.
 - `POST /predict_raw` accepts multipart `images[]` and optional normalized
   `box` and `yolo` controls. Each real result retains `filename`, top-64
   `leaves`, `wing_box`, `boxes`, and sex prediction. It also returns
-  `full_leaves`: all 7,841 unrounded leaf probabilities in head order, `leaf_distribution_complete: true`, and `pipeline_revision`.
+  `full_leaves`: all 7,482 unrounded leaf probabilities in head order, `leaf_distribution_complete: true`, and `pipeline_revision`.
   These are visual probabilities with **no geographic prior**. Mock results
   set `full_leaves: null` and `leaf_distribution_complete: false`.
 - `POST /predict` preserves the server-side geographic-prior response.
@@ -48,4 +49,4 @@ for full-distribution reranking.
 
 The Dockerfile pins model repository and BioCLIP revisions and verifies the
 head and detector bytes. The prior Space revision
-`cd88b0e612c0dc2a8c6c22647e5d15fe0140c7ea` is the rollback target.
+`fce1a2421dd87a753333c17b749af7749cec196b` is the rollback target.

inference.py

@@ -26,8 +26,8 @@ from PIL import Image, ImageOps
 TOP_K = 5
 RAW_TOP_N = 64   # how many leaves /predict_raw returns (client marginalises these)
 MAX_DETECTIONS = 32
-PIPELINE_REVISION = "wings-v6-gbif1000-head-20260927-v1"
-HEAD_SHA256 = "bc3c6275a69b6391b2520d53b267f011a9ab1405991dd8e2ab4be02b8b44c1ef"
+PIPELINE_REVISION = "wings-v6-gbif1000-vetted-taxonomy-20260927-v1"
+HEAD_SHA256 = "fe6f00821dc06cace5238d71c2561cc6eae37bcf1616129f65da8ccba5583eea"
 
 
 def deduplicate_detections(detections, image_size, max_groups=MAX_DETECTIONS):

test_release.py

@@ -66,7 +66,7 @@ class ReleaseTests(unittest.TestCase):
         self.assertEqual(hashlib.sha256(path.read_bytes()).hexdigest(), HEAD_SHA256)
         ck = torch.load(path, map_location="cpu", weights_only=False)
         n = len(ck["leaves"])
-        self.assertEqual(n, 7841)
+        self.assertGreater(n, 7000)
         self.assertEqual(ck["head_state_dict"]["weight"].shape[0], n)
         for rank in ("species", "genus"):
             self.assertEqual(len(ck["leaf2rank"][rank]), n)