diff --git a/tests/unit_tests/model_validation/sklearn/test_MinimumF1Score.py b/tests/unit_tests/model_validation/sklearn/test_MinimumF1Score.py
new file mode 100644
index 000000000..537609a6c
--- /dev/null
+++ b/tests/unit_tests/model_validation/sklearn/test_MinimumF1Score.py
@@ -0,0 +1,76 @@
+import unittest
+
+import numpy as np
+import pandas as pd
+from sklearn.linear_model import LogisticRegression
+from sklearn.metrics import f1_score
+
+import validmind as vm
+from validmind.tests.model_validation.sklearn.MinimumF1Score import MinimumF1Score
+
+
+def _dataset_with_predictions(input_id, y_true, y_pred):
+    """Build a VMDataset whose predictions are injected verbatim.
+
+    A fitted model is only needed to construct a valid VMModel; the predictions
+    are supplied via ``prediction_values`` so the model's ``predict`` is never
+    called and the true/predicted label sets can be controlled exactly.
+    """
+    df = pd.DataFrame(
+        {
+            "f1": np.linspace(-1.0, 1.0, len(y_true)),
+            "f2": np.linspace(1.0, -1.0, len(y_true)),
+            "target": y_true,
+        }
+    )
+    dataset = vm.init_dataset(
+        input_id=input_id, dataset=df, target_column="target", __log=False
+    )
+
+    model = LogisticRegression(max_iter=1000)
+    model.fit(df[["f1", "f2"]].to_numpy(), np.array(y_true))
+    vm_model = vm.init_model(input_id=f"{input_id}_model", model=model, __log=False)
+
+    dataset.assign_predictions(model=vm_model, prediction_values=y_pred)
+    return dataset, vm_model
+
+
+class TestMinimumF1Score(unittest.TestCase):
+    def test_predicted_class_absent_from_true_labels(self):
+        # Regression test for ZD-704. This split's true labels are binary ({0, 1}),
+        # but the model predicts a third class. scikit-learn's f1_score derives the
+        # target type from the union of y_true and y_pred, so deciding the averaging
+        # mode from y_true alone selects average="binary" and raises
+        # "Target is multiclass but average='binary'". The test must instead detect
+        # the multiclass label space and use macro averaging.
+        y_true = [0, 1, 1, 0, 1, 0, 1, 0]
+        y_pred = [0, 1, 2, 0, 1, 0, 2, 0]  # class 2 never appears in y_true
+        dataset, model = _dataset_with_predictions("f1_multiclass_pred", y_true, y_pred)
+
+        result = MinimumF1Score(dataset, model, min_threshold=0.5)
+
+        score = result[0][0]["Score"]
+        expected = f1_score(np.array(y_true), np.array(y_pred), average="macro")
+        # Matching the macro value confirms the multiclass branch was taken (the
+        # binary branch would have raised rather than returned a number).
+        self.assertAlmostEqual(score, expected)
+        self.assertEqual(result[1], score > 0.5)
+
+    def test_binary_uses_binary_average(self):
+        # A genuinely binary problem (true and predicted labels both within {0, 1})
+        # must still use sklearn's default binary averaging and be unaffected by the
+        # fix.
+        y_true = [0, 1, 1, 0, 1, 0]
+        y_pred = [0, 1, 0, 0, 1, 1]
+        dataset, model = _dataset_with_predictions("f1_binary", y_true, y_pred)
+
+        result = MinimumF1Score(dataset, model, min_threshold=0.5)
+
+        score = result[0][0]["Score"]
+        expected = f1_score(np.array(y_true), np.array(y_pred))  # binary default
+        self.assertAlmostEqual(score, expected)
+        self.assertEqual(result[1], score > 0.5)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/validmind/tests/model_validation/sklearn/MinimumF1Score.py b/validmind/tests/model_validation/sklearn/MinimumF1Score.py
index eaeb84bad..5deb83500 100644
--- a/validmind/tests/model_validation/sklearn/MinimumF1Score.py
+++ b/validmind/tests/model_validation/sklearn/MinimumF1Score.py
@@ -58,10 +58,19 @@ def MinimumF1Score(
     closely with specific requirements.
     """
 
-    if len(np.unique(dataset.y)) > 2:
-        score = f1_score(dataset.y, dataset.y_pred(model), average="macro")
+    y_true = dataset.y
+    y_pred = dataset.y_pred(model)
+
+    # Decide the averaging method from the labels f1_score actually sees -- the
+    # union of y_true and y_pred. Inspecting y_true alone is not enough: a split
+    # whose true labels collapse to <=2 classes while the model predicts a third
+    # class is still multiclass to sklearn, and average="binary" would raise
+    # "Target is multiclass but average='binary'".
+    n_classes = len(np.unique(np.concatenate([y_true, y_pred])))
+    if n_classes > 2:
+        score = f1_score(y_true, y_pred, average="macro")
     else:
-        score = f1_score(dataset.y, dataset.y_pred(model))
+        score = f1_score(y_true, y_pred)
 
     return (
         [
diff --git a/validmind/tests/model_validation/sklearn/PrecisionRecallCurve.py b/validmind/tests/model_validation/sklearn/PrecisionRecallCurve.py
index 321218404..c6b2bd76a 100644
--- a/validmind/tests/model_validation/sklearn/PrecisionRecallCurve.py
+++ b/validmind/tests/model_validation/sklearn/PrecisionRecallCurve.py
@@ -63,6 +63,7 @@ def PrecisionRecallCurve(
         raise SkipTestError("Skipping PrecisionRecallCurve for Foundation models")
 
     y_true = dataset.y
+    # Binary-only by design: multiclass is skipped, not handled (unlike MinimumF1Score).
     if len(np.unique(y_true)) > 2:
         raise SkipTestError(
             "Precision Recall Curve is only supported for binary classification models"
diff --git a/validmind/tests/model_validation/sklearn/ROCCurve.py b/validmind/tests/model_validation/sklearn/ROCCurve.py
index 7f8c99690..6109ca1d8 100644
--- a/validmind/tests/model_validation/sklearn/ROCCurve.py
+++ b/validmind/tests/model_validation/sklearn/ROCCurve.py
@@ -68,6 +68,7 @@ def ROCCurve(model: VMModel, dataset: VMDataset) -> Tuple[go.Figure, RawData]:
     incorrect, provided that the model's ranking format is retained. This phenomenon is commonly termed the "Class
     Imbalance Problem".
     """
+    # Binary-only by design: multiclass is skipped, not handled (unlike MinimumF1Score).
     if len(np.unique(dataset.y)) > 2:
         raise SkipTestError(
             "ROC Curve is only supported for binary classification models"