diff --git a/tensorflow_datasets/core/dataset_builder.py b/tensorflow_datasets/core/dataset_builder.py
index c4d0befd769..a964e03a720 100644
--- a/tensorflow_datasets/core/dataset_builder.py
+++ b/tensorflow_datasets/core/dataset_builder.py
@@ -890,18 +890,22 @@ def as_data_source(
         )
       chosen_format = file_format
     else:
-      chosen_format = suitable_formats.pop()
-      logging.info(
-          "Found random access formats: %s. Chose to use %s. Overriding file"
-          " format in the dataset info.",
-          ", ".join([f.name for f in suitable_formats]),
-          chosen_format,
-      )
+      if info.file_format in suitable_formats:
+        chosen_format = info.file_format
+      else:
+        chosen_format = suitable_formats.pop()
+        logging.info(
+            "Found random access formats: %s. Chose to use %s. Overriding file"
+            " format in the dataset info.",
+            ", ".join([f.name for f in suitable_formats]),
+            chosen_format,
+        )
 
-    # Change the dataset info to read from a random access format.
-    info.set_file_format(
-        chosen_format, override=True, override_if_initialized=True
-    )
+    if info.file_format != chosen_format:
+      # Change the dataset info to read from a random access format.
+      info.set_file_format(
+          chosen_format, override=True, override_if_initialized=True
+      )
 
     # Create a dataset for each of the given splits
     def build_single_data_source(split: str) -> Sequence[Any]:
diff --git a/tensorflow_datasets/core/dataset_builder_test.py b/tensorflow_datasets/core/dataset_builder_test.py
index d2d052c5c69..9c7d884f162 100644
--- a/tensorflow_datasets/core/dataset_builder_test.py
+++ b/tensorflow_datasets/core/dataset_builder_test.py
@@ -802,6 +802,31 @@ def test_load_as_data_source_alternative_file_format(self):
     assert len(data_source) == 10
     assert data_source[0]["x"] == 28
 
+  def test_load_as_data_source_with_multi_split_info(self):
+    data_dir = self.get_temp_dir()
+    builder = DummyDatasetWithConfigs(
+        data_dir=data_dir,
+        config="plus1",
+        file_format=file_adapters.FileFormat.ARRAY_RECORD,
+    )
+    builder.download_and_prepare()
+
+    # Make it a multi-split dataset.
+    multi_split_info = splits_lib.MultiSplitInfo(
+        name="train", split_infos=[builder.info.splits["train"]]
+    )
+    builder.info.set_splits(splits_lib.SplitDict([multi_split_info]))
+
+    self.assertIsInstance(
+        builder.info.splits["train"], splits_lib.MultiSplitInfo
+    )
+    self.assertIsNone(builder.info.splits["train"].filename_template)
+
+    data_source = builder.as_data_source(split="train")
+    self.assertIsNotNone(data_source)
+    self.assertLen(data_source, 20)
+    self.assertEqual(data_source[0]["x"], 7)
+
   @parameterized.named_parameters(
       *[
           {"file_format": file_format, "testcase_name": file_format.value}
diff --git a/tensorflow_datasets/core/dataset_info.py b/tensorflow_datasets/core/dataset_info.py
index 8fc52ab272f..f2eaf7ea0d5 100644
--- a/tensorflow_datasets/core/dataset_info.py
+++ b/tensorflow_datasets/core/dataset_info.py
@@ -538,6 +538,9 @@ def set_file_format(
       updated_split_infos = []
       for split_info in self.splits.values():
         if split_info.filename_template is None:
+          logging.warning(
+              "Split %s has no filename template, skipping.", split_info.name
+          )
           continue
         updated_split_info = split_info.replace(
             filename_template=split_info.filename_template.replace(