diff --git a/tests/python/conftest.py b/tests/python/conftest.py
new file mode 100644
index 00000000..e9bfae4c
--- /dev/null
+++ b/tests/python/conftest.py
@@ -0,0 +1,14 @@
+import multiprocessing
+import pytest
+
+
+# The default multiprocessing start method is "fork" which is not compatible with
+# with runtime assertions that it is set to spawn. When running unit tests, it's
+# possible to call an external library that sets the start method to "fork".
+# Here we enforce the start method to be "spawn" for all tests before executing.
+@pytest.fixture(scope="session", autouse=True)
+def initialize_data_environment():
+    try:
+        multiprocessing.set_start_method("spawn")
+    except Exception:
+        pass
diff --git a/tests/python/test_code_composition.py b/tests/python/test_code_composition.py
index f40bea49..5b7b5308 100644
--- a/tests/python/test_code_composition.py
+++ b/tests/python/test_code_composition.py
@@ -1,4 +1,3 @@
-import multiprocessing
 from unittest import TestCase
 
 from dolma.core.data_types import Document
@@ -40,11 +39,6 @@ def baz():
 
 class TestDolmaCodeProseCompositionClassifier(TestCase):
     def setUp(self) -> None:
-        try:
-            multiprocessing.set_start_method("spawn")
-        except Exception:
-            pass
-
         self.code_composition_tagger = CodeProseCompositionClassifier()
 
     def test_prose_text(self):