diff --git a/rasa/shared/nlu/training_data/formats/rasa_yaml.py b/rasa/shared/nlu/training_data/formats/rasa_yaml.py
index ce254b109b01..c41884b94386 100644
--- a/rasa/shared/nlu/training_data/formats/rasa_yaml.py
+++ b/rasa/shared/nlu/training_data/formats/rasa_yaml.py
@@ -8,6 +8,7 @@
 from rasa.shared.exceptions import YamlException
 from rasa.shared.utils import validation
 from ruamel.yaml import StringIO
+from ruamel.yaml.scalarstring import LiteralScalarString
 
 from rasa.shared.constants import (
     DOCS_URL_TRAINING_DATA,
@@ -36,6 +37,7 @@
 KEY_LOOKUP = "lookup"
 KEY_LOOKUP_EXAMPLES = "examples"
 KEY_METADATA = "metadata"
+KEY_METADATA_EXAMPLE = "example"
 
 MULTILINE_TRAINING_EXAMPLE_LEADING_SYMBOL = "-"
 
@@ -470,21 +472,57 @@ def process_training_examples_by_key(
         key_examples: Text,
         example_extraction_predicate=lambda x: x,
     ) -> List[OrderedDict]:
-        from ruamel.yaml.scalarstring import LiteralScalarString
-
         result = []
+
         for entity_key, examples in training_examples.items():
+            converted_examples = []
+            render_as_objects = False
+            for example in examples:
+                converted_example = {
+                    KEY_INTENT_TEXT: example_extraction_predicate(example)
+                }
 
-            converted_examples = [
-                TrainingDataWriter.generate_list_item(
-                    example_extraction_predicate(example).strip(STRIP_SYMBOLS)
-                )
-                for example in examples
-            ]
+                if isinstance(example, dict) and KEY_METADATA_EXAMPLE in example.get(
+                    KEY_METADATA, {}
+                ):
+                    render_as_objects = True
+                    converted_example[KEY_METADATA] = example[KEY_METADATA]["example"]
+
+                converted_examples.append(converted_example)
 
             next_item = OrderedDict()
             next_item[key_name] = entity_key
-            next_item[key_examples] = LiteralScalarString("".join(converted_examples))
+
+            if render_as_objects:
+                rendered_examples = RasaYAMLWriter._render_training_examples_as_objects(
+                    converted_examples
+                )
+            else:
+                rendered_examples = RasaYAMLWriter._render_training_examples_as_text(
+                    converted_examples
+                )
+            next_item[key_examples] = rendered_examples
+
             result.append(next_item)
 
         return result
+
+    @staticmethod
+    def _render_training_examples_as_objects(examples: List[Dict]) -> List[Dict]:
+        def render(example: Dict) -> Dict:
+            value = example[KEY_INTENT_TEXT]
+            example[KEY_INTENT_TEXT] = LiteralScalarString(
+                TrainingDataWriter.generate_string_item(value)
+            )
+            return example
+
+        return [render(ex) for ex in examples]
+
+    @staticmethod
+    def _render_training_examples_as_text(examples: List[Dict]) -> List[Text]:
+        def render(example: Dict) -> Text:
+            return TrainingDataWriter.generate_list_item(
+                example[KEY_INTENT_TEXT].strip(STRIP_SYMBOLS)
+            )
+
+        return LiteralScalarString("".join([render(example) for example in examples]))
diff --git a/rasa/shared/nlu/training_data/formats/readerwriter.py b/rasa/shared/nlu/training_data/formats/readerwriter.py
index 2a538833b929..5e2bba93fe89 100644
--- a/rasa/shared/nlu/training_data/formats/readerwriter.py
+++ b/rasa/shared/nlu/training_data/formats/readerwriter.py
@@ -69,8 +69,12 @@ def prepare_training_examples(training_data: "TrainingData") -> OrderedDict:
     @staticmethod
     def generate_list_item(text: Text) -> Text:
         """Generates text for a list item."""
+        return f"- {TrainingDataWriter.generate_string_item(text)}"
 
-        return f"- {rasa.shared.nlu.training_data.util.encode_string(text)}\n"
+    @staticmethod
+    def generate_string_item(text: Text) -> Text:
+        """Generates text for a string item."""
+        return f"{rasa.shared.nlu.training_data.util.encode_string(text)}\n"
 
     @staticmethod
     def generate_message(message: Dict[Text, Any]) -> Text:
diff --git a/tests/shared/nlu/training_data/formats/test_rasa_yaml.py b/tests/shared/nlu/training_data/formats/test_rasa_yaml.py
index 842e504b9ba4..383036b4d43f 100644
--- a/tests/shared/nlu/training_data/formats/test_rasa_yaml.py
+++ b/tests/shared/nlu/training_data/formats/test_rasa_yaml.py
@@ -66,6 +66,25 @@
     - Hello
 """
 
+INTENT_EXAMPLES_WITH_METADATA_ROUNDTRIP = f"""version: "{LATEST_TRAINING_DATA_FORMAT_VERSION}"
+nlu:
+- intent: intent_name
+  examples:
+  - text: |
+      how much CO2 will that use?
+    metadata:
+      sentiment: positive
+  - text: |
+      how much carbon will a one way flight from [new york]{{"entity": "city", "role": "from"}} to california produce?
+    metadata: co2-trip-calculation
+  - text: |
+      how much CO2 to [new york]{{"entity": "city", "role": "to"}}?
+- intent: greet
+  examples: |
+    - Hi
+    - Hello
+"""
+
 MINIMAL_VALID_EXAMPLE = """
 nlu:\n
 stories:
@@ -177,6 +196,22 @@ def test_intent_with_metadata_is_parsed():
     }
 
 
+def test_metadata_roundtrip():
+    reader = RasaYAMLReader()
+    result = reader.reads(INTENT_EXAMPLES_WITH_METADATA_ROUNDTRIP)
+
+    dumped = RasaYAMLWriter().dumps(result)
+    assert dumped == INTENT_EXAMPLES_WITH_METADATA_ROUNDTRIP
+
+    validation_reader = RasaYAMLReader()
+    dumped_result = validation_reader.reads(dumped)
+
+    assert dumped_result.training_examples == result.training_examples
+
+    # dumping again should also not change the format
+    assert dumped == RasaYAMLWriter().dumps(dumped_result)
+
+
 # This test would work only with examples that have a `version` key specified
 @pytest.mark.parametrize(
     "example",