Fix observation counts for unsupervised datasets

CHANGELOG.md

@@ -26,6 +26,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Allow JAX and JAXlib 0.11 in downstream environments by removing the
   `<0.11` dependency bounds ([#801](https://github.com/QuantClimate/GPJax/issues/801)).
 
+### Fixed
+
+- Read the observation count from `y` for unsupervised `Dataset` objects, so
+  `n` and `full_size` work when `X` is absent.
+
 ## [1.0.0] — 2026-09-28
 
 ### Added

gpjax/dataset.py

@@ -75,7 +75,7 @@ def __add__(self, other: "Dataset") -> "Dataset":
     @property
     def n(self) -> int:
         r"""Number of observations."""
-        return self.X.shape[0]
+        return self.X.shape[0] if self.X is not None else self.y.shape[0]
 
     @property
     def full_size(self) -> int:

tests/test_dataset.py

@@ -62,6 +62,17 @@ def test_dataset_init(n: int, in_dim: int) -> None:
     assert jtu.tree_leaves(D) == [x, y]
 
 
+@pytest.mark.parametrize("n", [0, 1, 3])
+@pytest.mark.parametrize("n_outputs", [1, 2])
+def test_unsupervised_dataset_observation_count(n, n_outputs):
+    data = Dataset(y=jnp.ones((n, n_outputs)))
+    assert data.is_unsupervised()
+    assert data.n == n
+    assert data.full_size == n
+    leaves, treedef = jtu.tree_flatten(data)
+    assert jtu.tree_unflatten(treedef, leaves).n == n
+
+
 @pytest.mark.parametrize("n1", [1, 2, 10])
 @pytest.mark.parametrize("n2", [1, 2, 10])
 @pytest.mark.parametrize("in_dim", [1, 2, 10])