fix: cap generalizability FF M_hat_mkpi to 1

Author: anyangml

lambench/metrics/utils.py

@@ -80,7 +80,9 @@ def filter_generalizability_force_field_results(
 
         if v is not None:
             if normalize:
-                v = v / std
+                v = np.min(
+                    [v / std, 1]
+                )  # cap the normalized value to 1, for models worese than a dummy baseline, use dummy baseline.
             filtered_metrics[k] = np.log(v) * weight
             # else the filtered_metrics will not have this key.
             # Metrics with weight != None should have a value,

tests/metrics/test_visualization.py

@@ -15,7 +15,7 @@ def test_aggregate_ood_results_for_one_model(
     aggregator = ResultsFetcher()
     result = aggregator.aggregate_ood_results_for_one_model(model=model)
     np.testing.assert_almost_equal(result["Small Molecules"], 0.19745455, decimal=5)
-    np.testing.assert_almost_equal(result["Inorganic Materials"], 0.3179398, decimal=5)
+    np.testing.assert_almost_equal(result["Inorganic Materials"], 0.283787, decimal=5)
     assert result["Catalysis"] is None
     with caplog.at_level(logging.WARNING):
         assert (