Create a new dataset object for each _generate_example call.

PiperOrigin-RevId: 716664290

Author: The TensorFlow Datasets Authors

tensorflow_datasets/core/dataset_builders/croissant_builder.py

@@ -217,15 +217,17 @@ def __init__(
     """
     if mapping is None:
       mapping = {}
-    self.dataset = mlc.Dataset(jsonld, mapping=mapping)
-    self.name = croissant_utils.get_tfds_dataset_name(self.dataset)
-    self.metadata = self.dataset.metadata
+    self.jsonld = jsonld
+    self.mapping = mapping
+    dataset = mlc.Dataset(jsonld, mapping=mapping)
+    self.name = croissant_utils.get_tfds_dataset_name(dataset)
+    self.metadata = dataset.metadata
 
     # In TFDS, version is a mandatory attribute, while in Croissant it is only a
     # recommended attribute. If the version is unspecified in Croissant, we set
     # it to `1.0.0` in TFDS.
     self.VERSION = version_lib.Version(  # pylint: disable=invalid-name
-        overwrite_version or self.dataset.metadata.version or '1.0.0'
+        overwrite_version or self.metadata.version or '1.0.0'
     )
     self.RELEASE_NOTES = {}  # pylint: disable=invalid-name
 
@@ -260,11 +262,11 @@ def builder_config(self) -> dataset_builder.BuilderConfig:
   def _info(self) -> dataset_info.DatasetInfo:
     return dataset_info.DatasetInfo(
         builder=self,
-        description=self.dataset.metadata.description,
+        description=self.metadata.description,
         features=self.get_features(),
-        homepage=self.dataset.metadata.url,
-        citation=self.dataset.metadata.cite_as,
-        license=_get_license(self.dataset.metadata),
+        homepage=self.metadata.url,
+        citation=self.metadata.cite_as,
+        license=_get_license(self.metadata),
         disable_shuffling=self._disable_shuffling,
     )
 
@@ -331,7 +333,8 @@ def _generate_examples(
     record_set = croissant_utils.get_record_set(
         self.builder_config.name, metadata=self.metadata
     )
-    records = self.dataset.records(record_set.id, filters=filters)
+    dataset = mlc.Dataset(self.jsonld, mapping=self.mapping)
+    records = dataset.records(record_set.id, filters=filters)
 
     def convert_to_tfds_format(
         global_index: int,

tensorflow_datasets/core/dataset_builders/croissant_builder_test.py

@@ -242,7 +242,10 @@ def test_croissant_builder(crs_builder):
 def test_download_and_prepare(crs_builder, expected_entries, split_name):
   crs_builder.download_and_prepare()
   data_source = crs_builder.as_data_source(split=split_name)
-  assert len(data_source) == 1
+  expected_entries = [
+      entry for entry in expected_entries if entry["split"] == split_name
+  ]
+  assert len(data_source) == len(expected_entries) == 1
   for entry, expected_entry in zip(data_source, expected_entries):
     assert entry["index"] == expected_entry["index"]
     assert entry["text"].decode() == expected_entry["text"]