diff --git a/data/satlas/wind_turbine/README.md b/data/satlas/wind_turbine/README.md
index 89ffce21..36274022 100644
--- a/data/satlas/wind_turbine/README.md
+++ b/data/satlas/wind_turbine/README.md
@@ -14,3 +14,12 @@ Dataset Versions
 - Sentinel-2 L1C (config.json): `rslearn-eai/datasets/wind_turbine/dataset_v1/20241002/`
 - Old Azure with Sentinel-2 L2A and Sentinel-1: `rslearn-eai/datasets/wind_turbine/dataset_v1/20241212/`
 - New Azure with Sentinel-2 L2A, Sentinel-1, and Landsat: `rslearn-eai/datasets/wind_turbine/dataset_v1/20250605/`
+- Fixed naip group: `rslearn-eai/datasets/wind_turbine/dataset_v1/20260122/`
+    - It looks like the coordinates of the labels in the naip group were not
+      re-projected properly when this dataset was first converted from the multisat
+      version -- they were still in WebMercator coordinates and so the model was
+      trained not to detect anything in all of the naip windows (which were originally
+      created by training and applying an object detector with high-res NAIP images,
+      and using those outputs to supervise Sentinel-2 training). So this new version is
+      the same as 20250605 but with the labels fixed. The script to fix the labels is
+      `one_off_projects/2026_01_22_fix_wind_turbine_dataset/run.py`.
diff --git a/one_off_projects/2026_01_22_fix_wind_turbine_dataset/run.py b/one_off_projects/2026_01_22_fix_wind_turbine_dataset/run.py
new file mode 100644
index 00000000..41a01ef8
--- /dev/null
+++ b/one_off_projects/2026_01_22_fix_wind_turbine_dataset/run.py
@@ -0,0 +1,79 @@
+"""Fix coordinates of labels in the naip group of the wind turbine dataset.
+
+Somehow these labels were re-projected incorrectly when the dataset was first converted
+from multisat format to rslearn dataset: although the "label" group seems okay, the
+"naip" group still has coordinates in WebMercator. This script fixes it. See
+data/satlas/wind_turbine/README.md for details.
+"""
+
+import argparse
+import json
+import math
+import multiprocessing
+
+import shapely
+import tqdm
+from rasterio.crs import CRS
+from rslearn.dataset import Dataset, Window
+from rslearn.utils.feature import Feature
+from rslearn.utils.geometry import Projection, STGeometry
+from rslearn.utils.vector_format import GeojsonVectorFormat, GeojsonCoordinateMode
+from upath import UPath
+
+# Compute the WebMercator projection for the source coordinates.
+SRC_CRS = CRS.from_epsg(3857)
+SRC_ZOOM = 13
+WEB_MERCATOR_M = 2 * math.pi * 6378137
+TOTAL_PIXELS_AT_ZOOM_LEVEL = (2 ** SRC_ZOOM) * 512
+SRC_PIXEL_SIZE = WEB_MERCATOR_M / TOTAL_PIXELS_AT_ZOOM_LEVEL
+SRC_PROJECTION = Projection(SRC_CRS, SRC_PIXEL_SIZE, -SRC_PIXEL_SIZE)
+
+DATASET_PATH = "/weka/dfive-default/rslearn-eai/datasets/wind_turbine/dataset_v1/20260122/"
+NUM_WORKERS = 64
+
+
+def fix_window(window: Window) -> None:
+    """Apply the fix to the specified window."""
+    label_dir = window.get_layer_dir("label")
+
+    # We read directly as JSON since the projection/coordinates aren't right, meaning
+    # that GeojsonVectorFormat won't work correctly.
+    with (label_dir / "data.geojson").open() as f:
+        fc = json.load(f)
+    if len(fc["features"]) == 0:
+        # We don't need to overwrite it since there are no features.
+        return
+
+    # Re-project each feature. We need to subtract TOTAL_PIXELS_AT_ZOOM_LEVEL because
+    # in rslearn the coordinate system starts from (-TOTAL_PIXELS_AT_ZOOM_LEVEL // 2),
+    # while in multisat it starts from 0.
+    new_features = []
+    for feat in fc["features"]:
+        col, row = feat["geometry"]["coordinates"]
+        src_geom = STGeometry(SRC_PROJECTION, shapely.Point(col - TOTAL_PIXELS_AT_ZOOM_LEVEL // 2, row - TOTAL_PIXELS_AT_ZOOM_LEVEL // 2), None)
+        dst_geom = src_geom.to_projection(window.projection)
+        new_feat = Feature(dst_geom, feat["properties"])
+
+        # Just double check that the resulting coordinates are reasonable. They might not
+        # be exactly contained within the window bounds, but it should be close.
+        window_geom = window.get_geometry()
+        distance = window_geom.shp.distance(dst_geom.shp)
+        if distance > 100:
+            raise ValueError(f"expected window {window.bounds} to contain re-projected geometry {dst_geom} but got distance {distance}")
+
+        new_features.append(new_feat)
+
+    # Now we can write out the GeoJSON. We use WGS84 so it is easy to verify in qgis.
+    vector_format = GeojsonVectorFormat(coordinate_mode=GeojsonCoordinateMode.WGS84)
+    vector_format.encode_vector(label_dir, new_features)
+
+
+if __name__ == "__main__":
+    multiprocessing.set_start_method("forkserver")
+    dataset = Dataset(UPath(DATASET_PATH))
+    windows = dataset.load_windows(groups=["naip"], workers=NUM_WORKERS, show_progress=True)
+    p = multiprocessing.Pool(128)
+    outputs = p.imap_unordered(fix_window, windows)
+    for _ in tqdm.tqdm(outputs, total=len(windows)):
+        pass
+    p.close()