diff --git a/src/metatrain/experimental/soap_bpnn/model.py b/src/metatrain/experimental/soap_bpnn/model.py
index a3d0d4ea..fc8e7e8c 100644
--- a/src/metatrain/experimental/soap_bpnn/model.py
+++ b/src/metatrain/experimental/soap_bpnn/model.py
@@ -123,14 +123,6 @@ def __init__(self, model_hypers: Dict, dataset_info: DatasetInfo) -> None:
             unit="unitless", per_atom=True
         )
 
-        # creates a composition weight tensor that can be directly indexed by species,
-        # this can be left as a tensor of zero or set from the outside using
-        # set_composition_weights (recommended for better accuracy)
-        n_outputs = len(self.outputs)
-        self.register_buffer(
-            "composition_weights",
-            torch.zeros((n_outputs, max(self.atomic_types) + 1)),
-        )
         # buffers cannot be indexed by strings (torchscript), so we create a single
         # tensor for all output. Due to this, we need to slice the tensor when we use
         # it and use the output name to select the correct slice via a dictionary
diff --git a/src/metatrain/experimental/soap_bpnn/tests/test_regression.py b/src/metatrain/experimental/soap_bpnn/tests/test_regression.py
index c36f337d..4845f240 100644
--- a/src/metatrain/experimental/soap_bpnn/tests/test_regression.py
+++ b/src/metatrain/experimental/soap_bpnn/tests/test_regression.py
@@ -41,10 +41,10 @@ def test_regression_init():
     expected_output = torch.tensor(
         [
             [-0.038599025458],
-            [ 0.111374437809],
-            [ 0.091115802526],
+            [0.111374437809],
+            [0.091115802526],
             [-0.056339077652],
-            [-0.025491207838]
+            [-0.025491207838],
         ]
     )
 
@@ -107,10 +107,10 @@ def test_regression_train():
     expected_output = torch.tensor(
         [
             [-0.106249026954],
-            [ 0.039981484413],
+            [0.039981484413],
             [-0.142682999372],
             [-0.031701669097],
-            [-0.016210660338]
+            [-0.016210660338],
         ]
     )