diff --git a/emm/indexing/pandas_normalized_tfidf.py b/emm/indexing/pandas_normalized_tfidf.py
index 4458615..3b3162b 100644
--- a/emm/indexing/pandas_normalized_tfidf.py
+++ b/emm/indexing/pandas_normalized_tfidf.py
@@ -37,8 +37,6 @@
 class PandasNormalizedTfidfVectorizer(TfidfVectorizer):
     """Implementation of customized TFIDF vectorizer"""
 
-    dtype = np.float32
-
     def __init__(self, **kwargs: Any) -> None:
         """Implementation of customized TFIDF vectorizer
 
@@ -53,6 +51,7 @@ def __init__(self, **kwargs: Any) -> None:
         Args:
             kwargs: kew-word arguments are same as TfidfVectorizer.
         """
+        kwargs.setdefault("dtype", np.float32)
         kwargs.update({"norm": None, "smooth_idf": True, "lowercase": True})
         if kwargs.get("analyzer") in {"word", None}:
             kwargs["token_pattern"] = r"\w+"
@@ -74,6 +73,8 @@ def fit(self, X: pd.Series | pd.DataFrame) -> TfidfVectorizer:
         with Timer("CustomizedTfidfVectorizer.fit") as timer:
             timer.label("super fit")
             super().fit(X)
+            # scikit-learn's TfidfVectorizer does not preserve dtype for large X, so we force it here
+            self.idf_ = self.idf_.astype(self.dtype)
 
             timer.label("normalize")
             n_features = self.idf_.shape[0]
diff --git a/emm/version.py b/emm/version.py
index 006f8eb..acb041b 100644
--- a/emm/version.py
+++ b/emm/version.py
@@ -17,6 +17,6 @@
 # IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 # CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 
-VERSION = "2.1.5"
+VERSION = "2.1.6"
 
 __version__ = VERSION
diff --git a/tests/integration/test_pandas_em.py b/tests/integration/test_pandas_em.py
index b60e132..f14e30d 100644
--- a/tests/integration/test_pandas_em.py
+++ b/tests/integration/test_pandas_em.py
@@ -21,6 +21,7 @@
 
 import logging
 import os
+import uuid
 
 import numpy as np
 import pandas as pd
@@ -135,6 +136,25 @@ def test_pandas_tfidf(dtype):
         np.testing.assert_allclose(actual_value, exp_value, rtol=0, atol=0.001)
 
 
+def test_pandas_tfidf_default_dtype():
+    pandas_t = PandasNormalizedTfidfVectorizer()
+    unique_names = [str(uuid.uuid4()) for i in range(100)]
+    gt_names = pd.Series(unique_names)
+    pandas_t.fit(gt_names)
+    assert pandas_t.idf_.dtype == np.float32
+
+
+@pytest.mark.parametrize(
+    ("dtype", "data_size"), [(np.float32, 100), (np.float64, 100), (np.float32, 1000000), (np.float64, 1000000)]
+)
+def test_pandas_tfidf_dtype_for_different_input_sizes(dtype, data_size):
+    pandas_t = PandasNormalizedTfidfVectorizer(dtype=dtype)
+    unique_names = [str(uuid.uuid4()) for i in range(data_size)]
+    gt_names = pd.Series(unique_names)
+    pandas_t.fit(gt_names)
+    assert pandas_t.idf_.dtype == dtype
+
+
 def test_pandas_tfidf_ngram():
     pandas_t = PandasNormalizedTfidfVectorizer(binary=True, analyzer="char", ngram_range=(3, 3))
     gt_names = pd.Series(["aaab", "bbbc"])