diff --git a/pyproject.toml b/pyproject.toml
index 7c8b8ae..35bf0a4 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -145,8 +145,8 @@ select = [
     "D210", "D211",
     # emptiness
     "D419",
-
-
+    # no pdb
+    "T100",
 ]
 ignore = [
     "ANN101", "ANN102", "ANN401", "ANN204",
diff --git a/src/numpydantic/interface/hdf5.py b/src/numpydantic/interface/hdf5.py
index 4fc548d..20cec0d 100644
--- a/src/numpydantic/interface/hdf5.py
+++ b/src/numpydantic/interface/hdf5.py
@@ -326,8 +326,16 @@ class H5Interface(Interface):
                     return np.datetime64
                 else:
                     return str
-            except (AttributeError, ValueError, TypeError):  # pragma: no cover
+            except (AttributeError, TypeError):  # pragma: no cover
+                # it's not a datetime, but it is some kind of string
                 return str
+            except (IndexError, ValueError):
+                # if the dataset is empty, we can't tell if something is a datetime
+                # or not, so we just tell the validation method what it wants to hear
+                if self.dtype in (np.datetime64, str):
+                    return self.dtype
+                else:
+                    return str
         else:
             return array.dtype
 
diff --git a/tests/test_interface/test_hdf5.py b/tests/test_interface/test_hdf5.py
index bf47a8d..9ca9e94 100644
--- a/tests/test_interface/test_hdf5.py
+++ b/tests/test_interface/test_hdf5.py
@@ -198,3 +198,23 @@ def test_datetime(hdf5_array, compound):
     assert instance.array[0, 0] == now
     instance.array[0] = now
     assert all(instance.array[0] == now)
+
+
+@pytest.mark.parametrize("dtype", [int, float, str, datetime])
+def test_empty_dataset(dtype, tmp_path):
+    """
+    Empty datasets shouldn't choke us during validation
+    """
+    array_path = tmp_path / "test.h5"
+    if dtype in (str, datetime):
+        np_dtype = "S32"
+    else:
+        np_dtype = dtype
+
+    with h5py.File(array_path, "w") as h5f:
+        _ = h5f.create_dataset(name="/data", dtype=np_dtype)
+
+    class MyModel(BaseModel):
+        array: NDArray[Any, dtype]
+
+    _ = MyModel(array=(array_path, "/data"))