diff --git a/changelog/4280.improvement.md b/changelog/4280.improvement.md
new file mode 100644
index 000000000000..542e9250bf69
--- /dev/null
+++ b/changelog/4280.improvement.md
@@ -0,0 +1 @@
+Improved CLI startup time.
diff --git a/changelog/8141.misc.md b/changelog/8141.misc.md
new file mode 100644
index 000000000000..c51cbfe1efe5
--- /dev/null
+++ b/changelog/8141.misc.md
@@ -0,0 +1,3 @@
+The following modules were renamed:
+* `rasa.train` -> `rasa.model_training`
+* `rasa.test` -> `rasa.model_testing`
diff --git a/changelog/8141.removal.md b/changelog/8141.removal.md
new file mode 100644
index 000000000000..5c4a2c581134
--- /dev/null
+++ b/changelog/8141.removal.md
@@ -0,0 +1,6 @@
+The following import abbreviations were removed:
+* `rasa.core.train`: Please use `rasa.core.train.train` instead.
+* `rasa.core.visualize`: Please use `rasa.core.visualize.visualize` instead.
+* `rasa.nlu.train`: Please use `rasa.nlu.train.train` instead.
+* `rasa.nlu.test`: Please use `rasa.nlu.test.run_evaluation` instead.
+* `rasa.nlu.cross_validate`: Please use `rasa.nlu.test.cross_validate` instead.
diff --git a/docs/docs/telemetry/telemetry.mdx b/docs/docs/telemetry/telemetry.mdx
index 0b36a52dcb8e..f4d532dabd87 100644
--- a/docs/docs/telemetry/telemetry.mdx
+++ b/docs/docs/telemetry/telemetry.mdx
@@ -113,7 +113,6 @@ Here is an example report that shows the data reported to Rasa after running
     "project": "a0a7178e6e5f9e6484c5cfa3ea4497ffc0c96d0ad3f3ad8e9399a1edd88e3cf4",
     "python": "3.7.5",
     "rasa_open_source": "2.0.0",
-    "gpu": 0,
     "cpu": 16
   }
 }
diff --git a/rasa/__init__.py b/rasa/__init__.py
index db331550b2c1..e10d5781c15d 100644
--- a/rasa/__init__.py
+++ b/rasa/__init__.py
@@ -1,12 +1,10 @@
 import logging
 
 from rasa import version
+from rasa.api import run, train, test
 
 # define the version before the other imports since these need it
 __version__ = version.__version__
 
-from rasa.run import run
-from rasa.train import train
-from rasa.test import test
 
 logging.getLogger(__name__).addHandler(logging.NullHandler())
diff --git a/rasa/api.py b/rasa/api.py
new file mode 100644
index 000000000000..a9d9ae10e49b
--- /dev/null
+++ b/rasa/api.py
@@ -0,0 +1,155 @@
+import rasa.shared.constants
+import typing
+
+# WARNING: Be careful about adding any top level imports at this place!
+#   These functions are imported in `rasa.__init__` and any top level import
+#   added here will get executed as soon as someone runs `import rasa`.
+#   Some imports are very slow (e.g. `tensorflow`) and we want them to get
+#   imported when running `import rasa`. If you add more imports here,
+#   please check that in the chain you are importing, no slow packages
+#   are getting imported.
+
+if typing.TYPE_CHECKING:
+    from typing import Any, Text, Dict, Union, List, Optional, NoReturn
+    from rasa.model_training import TrainingResult
+    import asyncio
+
+
+def run(
+    model: "Text",
+    endpoints: "Text",
+    connector: "Text" = None,
+    credentials: "Text" = None,
+    **kwargs: "Dict[Text, Any]",
+) -> "NoReturn":
+    """Runs a Rasa model.
+
+    Args:
+        model: Path to model archive.
+        endpoints: Path to endpoints file.
+        connector: Connector which should be use (overwrites `credentials`
+        field).
+        credentials: Path to channel credentials file.
+        **kwargs: Additional arguments which are passed to
+        `rasa.core.run.serve_application`.
+
+    """
+    import rasa.core.run
+    from rasa.core.utils import AvailableEndpoints
+    from rasa.shared.utils.cli import print_warning
+    import rasa.shared.utils.common
+    from rasa.shared.constants import DOCS_BASE_URL
+
+    _endpoints = AvailableEndpoints.read_endpoints(endpoints)
+
+    if not connector and not credentials:
+        connector = "rest"
+
+        print_warning(
+            f"No chat connector configured, falling back to the "
+            f"REST input channel. To connect your bot to another channel, "
+            f"read the docs here: {DOCS_BASE_URL}/messaging-and-voice-channels"
+        )
+
+    kwargs = rasa.shared.utils.common.minimal_kwargs(
+        kwargs, rasa.core.run.serve_application
+    )
+    rasa.core.run.serve_application(
+        model,
+        channel=connector,
+        credentials=credentials,
+        endpoints=_endpoints,
+        **kwargs,
+    )
+
+
+def train(
+    domain: "Text",
+    config: "Text",
+    training_files: "Union[Text, List[Text]]",
+    output: "Text" = rasa.shared.constants.DEFAULT_MODELS_PATH,
+    dry_run: bool = False,
+    force_training: bool = False,
+    fixed_model_name: "Optional[Text]" = None,
+    persist_nlu_training_data: bool = False,
+    core_additional_arguments: "Optional[Dict]" = None,
+    nlu_additional_arguments: "Optional[Dict]" = None,
+    loop: "Optional[asyncio.AbstractEventLoop]" = None,
+    model_to_finetune: "Optional[Text]" = None,
+    finetuning_epoch_fraction: float = 1.0,
+) -> "TrainingResult":
+    """Runs Rasa Core and NLU training in `async` loop.
+
+    Args:
+        domain: Path to the domain file.
+        config: Path to the config for Core and NLU.
+        training_files: Paths to the training data for Core and NLU.
+        output: Output path.
+        dry_run: If `True` then no training will be done, and the information about
+            whether the training needs to be done will be printed.
+        force_training: If `True` retrain model even if data has not changed.
+        fixed_model_name: Name of model to be stored.
+        persist_nlu_training_data: `True` if the NLU training data should be persisted
+            with the model.
+        core_additional_arguments: Additional training parameters for core training.
+        nlu_additional_arguments: Additional training parameters forwarded to training
+            method of each NLU component.
+        loop: Optional EventLoop for running coroutines.
+        model_to_finetune: Optional path to a model which should be finetuned or
+            a directory in case the latest trained model should be used.
+        finetuning_epoch_fraction: The fraction currently specified training epochs
+            in the model configuration which should be used for finetuning.
+
+    Returns:
+        An instance of `TrainingResult`.
+    """
+    from rasa.model_training import train_async
+    import rasa.utils.common
+
+    return rasa.utils.common.run_in_loop(
+        train_async(
+            domain=domain,
+            config=config,
+            training_files=training_files,
+            output=output,
+            dry_run=dry_run,
+            force_training=force_training,
+            fixed_model_name=fixed_model_name,
+            persist_nlu_training_data=persist_nlu_training_data,
+            core_additional_arguments=core_additional_arguments,
+            nlu_additional_arguments=nlu_additional_arguments,
+            model_to_finetune=model_to_finetune,
+            finetuning_epoch_fraction=finetuning_epoch_fraction,
+        ),
+        loop,
+    )
+
+
+def test(
+    model: "Text",
+    stories: "Text",
+    nlu_data: "Text",
+    output: "Text" = rasa.shared.constants.DEFAULT_RESULTS_PATH,
+    additional_arguments: "Optional[Dict]" = None,
+) -> None:
+    """Test a Rasa model against a set of test data.
+
+    Args:
+        model: model to test
+        stories: path to the dialogue test data
+        nlu_data: path to the NLU test data
+        output: path to folder where all output will be stored
+        additional_arguments: additional arguments for the test call
+    """
+    from rasa.model_testing import test_core
+    import rasa.utils.common
+    from rasa.model_testing import test_nlu
+
+    if additional_arguments is None:
+        additional_arguments = {}
+
+    test_core(model, stories, output, additional_arguments)
+
+    rasa.utils.common.run_in_loop(
+        test_nlu(model, nlu_data, output, additional_arguments)
+    )
diff --git a/rasa/cli/data.py b/rasa/cli/data.py
index 00f7752e2050..0fd3abb8f6c3 100644
--- a/rasa/cli/data.py
+++ b/rasa/cli/data.py
@@ -11,11 +11,6 @@
 from rasa.cli.arguments import data as arguments
 from rasa.cli.arguments import default_arguments
 import rasa.cli.utils
-from rasa.core.training.converters.responses_prefix_converter import (
-    DomainResponsePrefixConverter,
-    StoryResponsePrefixConverter,
-)
-import rasa.nlu.convert
 from rasa.shared.constants import (
     DEFAULT_DATA_PATH,
     DEFAULT_CONFIG_PATH,
@@ -24,6 +19,10 @@
 )
 import rasa.shared.data
 from rasa.shared.core.constants import (
+    POLICY_NAME_FALLBACK,
+    POLICY_NAME_FORM,
+    POLICY_NAME_MAPPING,
+    POLICY_NAME_TWO_STAGE_FALLBACK,
     USER_INTENT_OUT_OF_SCOPE,
     ACTION_DEFAULT_FALLBACK_NAME,
 )
@@ -38,18 +37,13 @@
 import rasa.shared.nlu.training_data.util
 import rasa.shared.utils.cli
 import rasa.utils.common
-from rasa.utils.converter import TrainingDataConverter
-from rasa.validator import Validator
 from rasa.shared.core.domain import Domain, InvalidDomain
 import rasa.shared.utils.io
-import rasa.core.config
-from rasa.core.policies.form_policy import FormPolicy
-from rasa.core.policies.fallback import FallbackPolicy
-from rasa.core.policies.two_stage_fallback import TwoStageFallbackPolicy
-from rasa.core.policies.mapping_policy import MappingPolicy
 
 if TYPE_CHECKING:
     from rasa.shared.core.training_data.structures import StoryStep
+    from rasa.validator import Validator
+    from rasa.utils.converter import TrainingDataConverter
 
 logger = logging.getLogger(__name__)
 
@@ -242,6 +236,7 @@ def validate_files(args: argparse.Namespace, stories_only: bool = False) -> None
         args: Commandline arguments
         stories_only: If `True`, only the story structure is validated.
     """
+    from rasa.validator import Validator
 
     config = rasa.cli.utils.get_validated_path(
         args.config, "config", DEFAULT_CONFIG_PATH, none_is_valid=True
@@ -278,15 +273,15 @@ def validate_stories(args: argparse.Namespace) -> None:
     validate_files(args, stories_only=True)
 
 
-def _validate_domain(validator: Validator) -> bool:
+def _validate_domain(validator: "Validator") -> bool:
     return validator.verify_domain_validity()
 
 
-def _validate_nlu(validator: Validator, args: argparse.Namespace) -> bool:
+def _validate_nlu(validator: "Validator", args: argparse.Namespace) -> bool:
     return validator.verify_nlu(not args.fail_on_warnings)
 
 
-def _validate_story_structure(validator: Validator, args: argparse.Namespace) -> bool:
+def _validate_story_structure(validator: "Validator", args: argparse.Namespace) -> bool:
     # Check if a valid setting for `max_history` was given
     if isinstance(args.max_history, int) and args.max_history < 1:
         raise argparse.ArgumentTypeError(
@@ -299,6 +294,8 @@ def _validate_story_structure(validator: Validator, args: argparse.Namespace) ->
 
 
 def _convert_nlu_data(args: argparse.Namespace) -> None:
+    import rasa.nlu.convert
+
     from rasa.nlu.training_data.converters.nlu_markdown_to_yaml_converter import (
         NLUMarkdownToYamlConverter,
     )
@@ -356,8 +353,14 @@ def _convert_nlg_data(args: argparse.Namespace) -> None:
 
 def _migrate_responses(args: argparse.Namespace) -> None:
     """Migrate retrieval intent responses to the new 2.0 format.
+
     It does so modifying the stories and domain files.
     """
+    from rasa.core.training.converters.responses_prefix_converter import (
+        DomainResponsePrefixConverter,
+        StoryResponsePrefixConverter,
+    )
+
     if args.format == "yaml":
         rasa.utils.common.run_in_loop(
             _convert_to_yaml(args.out, args.domain, DomainResponsePrefixConverter())
@@ -374,7 +377,7 @@ def _migrate_responses(args: argparse.Namespace) -> None:
 
 
 async def _convert_to_yaml(
-    out_path: Text, data_path: Text, converter: TrainingDataConverter
+    out_path: Text, data_path: Text, converter: "TrainingDataConverter"
 ) -> None:
 
     output = Path(out_path)
@@ -415,7 +418,7 @@ async def _convert_to_yaml(
 
 
 async def _convert_file_to_yaml(
-    source_file: Path, target_dir: Path, converter: TrainingDataConverter
+    source_file: Path, target_dir: Path, converter: "TrainingDataConverter"
 ) -> bool:
     """Converts a single training data file to `YAML` format.
 
@@ -447,6 +450,8 @@ def _migrate_model_config(args: argparse.Namespace) -> None:
     Args:
         args: The commandline args with the required paths.
     """
+    import rasa.core.config
+
     configuration_file = Path(args.config)
     model_configuration = _get_configuration(configuration_file)
 
@@ -499,7 +504,7 @@ def _get_configuration(path: Path) -> Dict:
     _assert_two_stage_fallback_policy_is_migratable(config)
     _assert_only_one_fallback_policy_present(policy_names)
 
-    if FormPolicy.__name__ in policy_names:
+    if POLICY_NAME_FORM in policy_names:
         _warn_about_manual_forms_migration()
 
     return config
@@ -507,24 +512,24 @@ def _get_configuration(path: Path) -> Dict:
 
 def _assert_config_needs_migration(policies: List[Text]) -> None:
     migratable_policies = {
-        MappingPolicy.__name__,
-        FallbackPolicy.__name__,
-        TwoStageFallbackPolicy.__name__,
+        POLICY_NAME_MAPPING,
+        POLICY_NAME_FALLBACK,
+        POLICY_NAME_TWO_STAGE_FALLBACK,
     }
 
     if not migratable_policies.intersection((set(policies))):
         rasa.shared.utils.cli.print_error_and_exit(
             f"No policies were found which need migration. This command can migrate "
-            f"'{MappingPolicy.__name__}', '{FallbackPolicy.__name__}' and "
-            f"'{TwoStageFallbackPolicy.__name__}'."
+            f"'{POLICY_NAME_MAPPING}', '{POLICY_NAME_FALLBACK}' and "
+            f"'{POLICY_NAME_TWO_STAGE_FALLBACK}'."
         )
 
 
 def _warn_about_manual_forms_migration() -> None:
     rasa.shared.utils.cli.print_warning(
-        f"Your model configuration contains the '{FormPolicy.__name__}'. "
-        f"Note that this command does not migrate the '{FormPolicy.__name__}' and "
-        f"you have to migrate the '{FormPolicy.__name__}' manually. "
+        f"Your model configuration contains the '{POLICY_NAME_FORM}'. "
+        f"Note that this command does not migrate the '{POLICY_NAME_FORM}' and "
+        f"you have to migrate the '{POLICY_NAME_FORM}' manually. "
         f"Please see the migration guide for further details: "
         f"{DOCS_URL_MIGRATION_GUIDE}"
     )
@@ -533,7 +538,7 @@ def _warn_about_manual_forms_migration() -> None:
 def _assert_nlu_pipeline_given(config: Dict, policy_names: List[Text]) -> None:
     if not config.get("pipeline") and any(
         policy in policy_names
-        for policy in [FallbackPolicy.__name__, TwoStageFallbackPolicy.__name__]
+        for policy in [POLICY_NAME_FALLBACK, POLICY_NAME_TWO_STAGE_FALLBACK]
     ):
         rasa.shared.utils.cli.print_error_and_exit(
             "The model configuration has to include an NLU pipeline. This is required "
@@ -546,7 +551,7 @@ def _assert_two_stage_fallback_policy_is_migratable(config: Dict) -> None:
         (
             policy_config
             for policy_config in config.get("policies", [])
-            if policy_config.get("name") == TwoStageFallbackPolicy.__name__
+            if policy_config.get("name") == POLICY_NAME_TWO_STAGE_FALLBACK
         ),
         None,
     )
@@ -583,10 +588,7 @@ def _assert_two_stage_fallback_policy_is_migratable(config: Dict) -> None:
 
 
 def _assert_only_one_fallback_policy_present(policies: List[Text]) -> None:
-    if (
-        FallbackPolicy.__name__ in policies
-        and TwoStageFallbackPolicy.__name__ in policies
-    ):
+    if POLICY_NAME_FALLBACK in policies and POLICY_NAME_TWO_STAGE_FALLBACK in policies:
         rasa.shared.utils.cli.print_error_and_exit(
             "Your policy configuration contains two configured policies for handling "
             "fallbacks. Please decide on one."
diff --git a/rasa/cli/run.py b/rasa/cli/run.py
index 99c785b1d607..aa964da88b0c 100644
--- a/rasa/cli/run.py
+++ b/rasa/cli/run.py
@@ -1,7 +1,7 @@
 import argparse
 import logging
 import os
-from typing import List, Text
+from typing import List, Text, NoReturn
 
 from rasa.cli import SubParsersAction
 from rasa.cli.arguments import run as arguments
@@ -74,8 +74,13 @@ def _validate_model_path(model_path: Text, parameter: Text, default: Text):
     return model_path
 
 
-def run(args: argparse.Namespace):
-    import rasa.run
+def run(args: argparse.Namespace) -> NoReturn:
+    """Entrypoint for `rasa run`.
+
+    Args:
+        args: The CLI arguments.
+    """
+    import rasa
 
     args.endpoints = rasa.cli.utils.get_validated_path(
         args.endpoints, "endpoints", DEFAULT_ENDPOINTS_PATH, True
diff --git a/rasa/cli/scaffold.py b/rasa/cli/scaffold.py
index 5357b8702463..fd5f30e9f673 100644
--- a/rasa/cli/scaffold.py
+++ b/rasa/cli/scaffold.py
@@ -5,7 +5,6 @@
 
 from rasa import telemetry
 from rasa.cli import SubParsersAction
-import rasa.train
 from rasa.cli.shell import shell
 from rasa.cli.utils import create_output_path
 from rasa.shared.utils.cli import print_success, print_error_and_exit
@@ -48,7 +47,9 @@ def add_subparser(
 
 
 def print_train_or_instructions(args: argparse.Namespace, path: Text) -> None:
+    """Train a model if the user wants to."""
     import questionary
+    import rasa
 
     print_success("Finished creating project structure.")
 
diff --git a/rasa/cli/test.py b/rasa/cli/test.py
index 4218ff3f81fd..a98847d9cfc1 100644
--- a/rasa/cli/test.py
+++ b/rasa/cli/test.py
@@ -17,7 +17,6 @@
     DEFAULT_DATA_PATH,
     DEFAULT_RESULTS_PATH,
 )
-from rasa.core.test import FAILED_STORIES_FILE
 import rasa.shared.utils.validation as validation_utils
 import rasa.cli.utils
 import rasa.utils.common
@@ -69,7 +68,12 @@ def add_subparser(
 
 def run_core_test(args: argparse.Namespace) -> None:
     """Run core tests."""
-    from rasa.test import test_core_models_in_directory, test_core, test_core_models
+    from rasa.model_testing import (
+        test_core_models_in_directory,
+        test_core,
+        test_core_models,
+    )
+    from rasa.core.test import FAILED_STORIES_FILE
 
     stories = rasa.cli.utils.get_validated_path(
         args.stories, "stories", DEFAULT_DATA_PATH
@@ -144,7 +148,11 @@ async def run_nlu_test_async(
         no_errors: indicates if incorrect predictions should be written to a file
                    or not.
     """
-    from rasa.test import compare_nlu_models, perform_nlu_cross_validation, test_nlu
+    from rasa.model_testing import (
+        compare_nlu_models,
+        perform_nlu_cross_validation,
+        test_nlu,
+    )
 
     nlu_data = rasa.cli.utils.get_validated_path(data_path, "nlu", DEFAULT_DATA_PATH)
     nlu_data = rasa.shared.data.get_nlu_directory(nlu_data)
diff --git a/rasa/cli/train.py b/rasa/cli/train.py
index 7f8cb4f8f2b8..c0c14d989810 100644
--- a/rasa/cli/train.py
+++ b/rasa/cli/train.py
@@ -130,7 +130,7 @@ def run_core_training(
     Returns:
         Path to a trained model or `None` if training was not successful.
     """
-    from rasa.train import train_core
+    from rasa.model_training import train_core
 
     output = train_path or args.out
 
@@ -179,7 +179,7 @@ def run_nlu_training(
     Returns:
         Path to a trained model or `None` if training was not successful.
     """
-    from rasa.train import train_nlu
+    from rasa.model_training import train_nlu
 
     output = train_path or args.out
 
diff --git a/rasa/cli/visualize.py b/rasa/cli/visualize.py
index 17b2ebff7770..0a69fc8929eb 100644
--- a/rasa/cli/visualize.py
+++ b/rasa/cli/visualize.py
@@ -38,7 +38,7 @@ def visualize_stories(args: argparse.Namespace):
         args.nlu = rasa.shared.data.get_nlu_directory(DEFAULT_DATA_PATH)
 
     rasa.utils.common.run_in_loop(
-        rasa.core.visualize(
+        rasa.core.visualize.visualize(
             args.config, args.domain, args.stories, args.nlu, args.out, args.max_history
         )
     )
diff --git a/rasa/core/__init__.py b/rasa/core/__init__.py
index 0f76ff881aef..4defa9baa4c2 100644
--- a/rasa/core/__init__.py
+++ b/rasa/core/__init__.py
@@ -2,9 +2,6 @@
 
 import rasa
 
-from rasa.core.train import train
-from rasa.core.visualize import visualize
-
 logging.getLogger(__name__).addHandler(logging.NullHandler())
 
 __version__ = rasa.__version__
diff --git a/rasa/core/agent.py b/rasa/core/agent.py
index fa3fe0d17e00..ae3784a4101a 100644
--- a/rasa/core/agent.py
+++ b/rasa/core/agent.py
@@ -11,6 +11,7 @@
 from aiohttp import ClientError
 
 import rasa
+import rasa.utils
 from rasa.core import jobs, training
 from rasa.core.channels.channel import OutputChannel, UserMessage
 from rasa.core.constants import DEFAULT_REQUEST_TIMEOUT
@@ -269,6 +270,33 @@ async def schedule_model_pulling(
     )
 
 
+def create_agent(model: Text, endpoints: Text = None) -> "Agent":
+    """Create an agent instance based on a stored model.
+
+    Args:
+        model: file path to the stored model
+        endpoints: file path to the used endpoint configuration
+    """
+    from rasa.core.tracker_store import TrackerStore
+    from rasa.core.utils import AvailableEndpoints
+    from rasa.core.brokers.broker import EventBroker
+    import rasa.utils.common
+
+    _endpoints = AvailableEndpoints.read_endpoints(endpoints)
+
+    _broker = rasa.utils.common.run_in_loop(EventBroker.create(_endpoints.event_broker))
+    _tracker_store = TrackerStore.create(_endpoints.tracker_store, event_broker=_broker)
+    _lock_store = LockStore.create(_endpoints.lock_store)
+
+    return Agent.load(
+        model,
+        generator=_endpoints.nlg,
+        tracker_store=_tracker_store,
+        lock_store=_lock_store,
+        action_endpoint=_endpoints.action,
+    )
+
+
 async def load_agent(
     model_path: Optional[Text] = None,
     model_server: Optional[EndpointConfig] = None,
@@ -278,7 +306,23 @@ async def load_agent(
     tracker_store: Optional[TrackerStore] = None,
     lock_store: Optional[LockStore] = None,
     action_endpoint: Optional[EndpointConfig] = None,
-):
+) -> Optional["Agent"]:
+    """Loads agent from server, remote storage or disk.
+
+    Args:
+        model_path: Path to the model if it's on disk.
+        model_server: Configuration for a potential server which serves the model.
+        remote_storage: URL of remote storage for model.
+        interpreter: NLU interpreter to parse incoming messages.
+        generator: Optional response generator.
+        tracker_store: TrackerStore for persisting the conversation history.
+        lock_store: LockStore to avoid that a conversation is modified by concurrent
+            actors.
+        action_endpoint: Action server configuration for executing custom actions.
+
+    Returns:
+        The instantiated `Agent` or `None`.
+    """
     try:
         if model_server is not None:
             return await load_from_server(
diff --git a/rasa/core/config.py b/rasa/core/config.py
index 2080a5dde667..a0936541a1c7 100644
--- a/rasa/core/config.py
+++ b/rasa/core/config.py
@@ -13,6 +13,11 @@
 from rasa.shared.core.constants import (
     ACTION_DEFAULT_FALLBACK_NAME,
     ACTION_TWO_STAGE_FALLBACK_NAME,
+    POLICY_NAME_RULE,
+    POLICY_NAME_FALLBACK,
+    POLICY_NAME_MAPPING,
+    POLICY_NAME_TWO_STAGE_FALLBACK,
+    CLASSIFIER_NAME_FALLBACK,
 )
 import rasa.utils.io
 from rasa.shared.constants import (
@@ -25,11 +30,6 @@
 
 import rasa.shared.utils.io
 import rasa.utils.io
-from rasa.core.policies.mapping_policy import MappingPolicy
-from rasa.core.policies.rule_policy import RulePolicy
-from rasa.core.policies.fallback import FallbackPolicy
-from rasa.core.policies.two_stage_fallback import TwoStageFallbackPolicy
-from rasa.nlu.classifiers.fallback_classifier import FallbackClassifier
 
 if TYPE_CHECKING:
     from rasa.core.policies.policy import Policy
@@ -69,8 +69,8 @@ def migrate_fallback_policies(config: Dict) -> Tuple[Dict, Optional["StoryStep"]
     policies = new_config.get("policies", [])
 
     fallback_config = _get_config_for_name(
-        FallbackPolicy.__name__, policies
-    ) or _get_config_for_name(TwoStageFallbackPolicy.__name__, policies)
+        POLICY_NAME_FALLBACK, policies
+    ) or _get_config_for_name(POLICY_NAME_TWO_STAGE_FALLBACK, policies)
 
     if not fallback_config:
         return config, None
@@ -83,7 +83,7 @@ def migrate_fallback_policies(config: Dict) -> Tuple[Dict, Optional["StoryStep"]
 
     # The triggered action is hardcoded for the Two-Stage Fallback`
     fallback_action_name = ACTION_TWO_STAGE_FALLBACK_NAME
-    if fallback_config.get("name") == FallbackPolicy.__name__:
+    if fallback_config.get("name") == POLICY_NAME_FALLBACK:
         fallback_action_name = fallback_config.get(
             "fallback_action_name", ACTION_DEFAULT_FALLBACK_NAME
         )
@@ -113,10 +113,10 @@ def _update_rule_policy_config_for_fallback(
         policies: The current list of configured policies.
         fallback_config: The configuration of the deprecated fallback configuration.
     """
-    rule_policy_config = _get_config_for_name(RulePolicy.__name__, policies)
+    rule_policy_config = _get_config_for_name(POLICY_NAME_RULE, policies)
 
     if not rule_policy_config:
-        rule_policy_config = {"name": RulePolicy.__name__}
+        rule_policy_config = {"name": POLICY_NAME_RULE}
         policies.append(rule_policy_config)
 
     core_threshold = fallback_config.get(
@@ -132,11 +132,11 @@ def _update_rule_policy_config_for_fallback(
 
 def _update_fallback_config(config: Dict, fallback_config: Dict) -> None:
     fallback_classifier_config = _get_config_for_name(
-        FallbackClassifier.__name__, config.get("pipeline", [])
+        CLASSIFIER_NAME_FALLBACK, config.get("pipeline", [])
     )
 
     if not fallback_classifier_config:
-        fallback_classifier_config = {"name": FallbackClassifier.__name__}
+        fallback_classifier_config = {"name": CLASSIFIER_NAME_FALLBACK}
         config["pipeline"].append(fallback_classifier_config)
 
     nlu_threshold = fallback_config.get("nlu_threshold", DEFAULT_NLU_FALLBACK_THRESHOLD)
@@ -188,15 +188,15 @@ def migrate_mapping_policy_to_rules(
     has_rule_policy = False
 
     for policy in policies:
-        if policy.get("name") == MappingPolicy.__name__:
+        if policy.get("name") == POLICY_NAME_MAPPING:
             has_mapping_policy = True
-        if policy.get("name") == RulePolicy.__name__:
+        if policy.get("name") == POLICY_NAME_RULE:
             has_rule_policy = True
 
     if not has_mapping_policy:
         return config, domain, []
 
-    rasa.shared.utils.cli.print_info(f"Migrating the '{MappingPolicy.__name__}'.")
+    rasa.shared.utils.cli.print_info(f"Migrating the '{POLICY_NAME_MAPPING}'.")
     new_config = copy.deepcopy(config)
     new_domain = copy.deepcopy(domain)
 
@@ -214,10 +214,10 @@ def migrate_mapping_policy_to_rules(
             new_rules.append(trigger_rule)
 
     # finally update the policies
-    policies = _drop_policy(MappingPolicy.__name__, policies)
+    policies = _drop_policy(POLICY_NAME_MAPPING, policies)
 
     if not has_rule_policy:
-        policies.append({"name": RulePolicy.__name__})
+        policies.append({"name": POLICY_NAME_RULE})
     new_config["policies"] = policies
 
     return new_config, new_domain, new_rules
diff --git a/rasa/core/test.py b/rasa/core/test.py
index efc60300fd50..adaa58160585 100644
--- a/rasa/core/test.py
+++ b/rasa/core/test.py
@@ -832,7 +832,7 @@ async def test(
     Returns:
         Evaluation summary.
     """
-    from rasa.test import get_evaluation_metrics
+    from rasa.model_testing import get_evaluation_metrics
 
     generator = await _create_data_generator(stories, agent, max_stories, e2e)
     completed_trackers = generator.generate_story_trackers()
diff --git a/rasa/core/training/interactive.py b/rasa/core/training/interactive.py
index 2637fdaa17e2..e399cab3ba50 100644
--- a/rasa/core/training/interactive.py
+++ b/rasa/core/training/interactive.py
@@ -34,7 +34,8 @@
     REQUESTED_SLOT,
     LOOP_INTERRUPTED,
 )
-from rasa.core import run, train, utils
+from rasa.core import run, utils
+import rasa.core.train
 from rasa.core.constants import DEFAULT_SERVER_FORMAT, DEFAULT_SERVER_PORT
 from rasa.shared.core.domain import Domain
 import rasa.shared.core.events
@@ -1628,7 +1629,7 @@ async def train_agent_on_start(
 
     model_directory = args.get("out", tempfile.mkdtemp(suffix="_core_model"))
 
-    _agent = await train(
+    _agent = await rasa.core.train.train(
         args.get("domain"),
         args.get("stories"),
         model_directory,
diff --git a/rasa/jupyter.py b/rasa/jupyter.py
index e57bda6d0cee..a0d516bc4640 100644
--- a/rasa/jupyter.py
+++ b/rasa/jupyter.py
@@ -8,10 +8,11 @@
 import rasa.utils.common
 
 if typing.TYPE_CHECKING:
-    from rasa.core.agent import Agent
+    from rasa.core.agent import Agent, create_agent
 
 
-def pprint(obj: Any):
+def pprint(obj: Any) -> None:
+    """Prints JSONs with indent."""
     pretty_print.pprint(obj, indent=2)
 
 
@@ -32,7 +33,6 @@ def chat(
     """
 
     if model_path:
-        from rasa.run import create_agent
 
         agent = create_agent(model_path, endpoints)
 
diff --git a/rasa/test.py b/rasa/model_testing.py
similarity index 96%
rename from rasa/test.py
rename to rasa/model_testing.py
index db3a8524ee10..ad954ed951a9 100644
--- a/rasa/test.py
+++ b/rasa/model_testing.py
@@ -90,20 +90,8 @@ def test_core_models(models: List[Text], stories: Text, output: Text):
     rasa.utils.common.run_in_loop(compare_models(models, stories, output))
 
 
-def test(
-    model: Text,
-    stories: Text,
-    nlu_data: Text,
-    output: Text = DEFAULT_RESULTS_PATH,
-    additional_arguments: Optional[Dict] = None,
-):
-    if additional_arguments is None:
-        additional_arguments = {}
-
-    test_core(model, stories, output, additional_arguments)
-    rasa.utils.common.run_in_loop(
-        test_nlu(model, nlu_data, output, additional_arguments)
-    )
+# backwards compatibility
+test = rasa.test
 
 
 def test_core(
diff --git a/rasa/train.py b/rasa/model_training.py
similarity index 92%
rename from rasa/train.py
rename to rasa/model_training.py
index e87136702c1c..b8106fe365df 100644
--- a/rasa/train.py
+++ b/rasa/model_training.py
@@ -51,63 +51,8 @@ class TrainingResult(NamedTuple):
     code: int = 0
 
 
-def train(
-    domain: Text,
-    config: Text,
-    training_files: Union[Text, List[Text]],
-    output: Text = DEFAULT_MODELS_PATH,
-    dry_run: bool = False,
-    force_training: bool = False,
-    fixed_model_name: Optional[Text] = None,
-    persist_nlu_training_data: bool = False,
-    core_additional_arguments: Optional[Dict] = None,
-    nlu_additional_arguments: Optional[Dict] = None,
-    loop: Optional[asyncio.AbstractEventLoop] = None,
-    model_to_finetune: Optional[Text] = None,
-    finetuning_epoch_fraction: float = 1.0,
-) -> TrainingResult:
-    """Runs Rasa Core and NLU training in `async` loop.
-
-    Args:
-        domain: Path to the domain file.
-        config: Path to the config for Core and NLU.
-        training_files: Paths to the training data for Core and NLU.
-        output: Output path.
-        dry_run: If `True` then no training will be done, and the information about
-            whether the training needs to be done will be printed.
-        force_training: If `True` retrain model even if data has not changed.
-        fixed_model_name: Name of model to be stored.
-        persist_nlu_training_data: `True` if the NLU training data should be persisted
-            with the model.
-        core_additional_arguments: Additional training parameters for core training.
-        nlu_additional_arguments: Additional training parameters forwarded to training
-            method of each NLU component.
-        loop: Optional EventLoop for running coroutines.
-        model_to_finetune: Optional path to a model which should be finetuned or
-            a directory in case the latest trained model should be used.
-        finetuning_epoch_fraction: The fraction currently specified training epochs
-            in the model configuration which should be used for finetuning.
-
-    Returns:
-        An instance of `TrainingResult`.
-    """
-    return rasa.utils.common.run_in_loop(
-        train_async(
-            domain=domain,
-            config=config,
-            training_files=training_files,
-            output=output,
-            dry_run=dry_run,
-            force_training=force_training,
-            fixed_model_name=fixed_model_name,
-            persist_nlu_training_data=persist_nlu_training_data,
-            core_additional_arguments=core_additional_arguments,
-            nlu_additional_arguments=nlu_additional_arguments,
-            model_to_finetune=model_to_finetune,
-            finetuning_epoch_fraction=finetuning_epoch_fraction,
-        ),
-        loop,
-    )
+# backwards compatibility
+train = rasa.train
 
 
 async def train_async(
@@ -600,7 +545,7 @@ async def _train_core_with_validated_data(
             model_type="core",
             is_finetuning=model_to_finetune is not None,
         ):
-            await rasa.core.train(
+            await rasa.core.train.train(
                 domain_file=domain,
                 training_resource=file_importer,
                 output_path=os.path.join(_train_path, DEFAULT_CORE_SUBDIRECTORY_NAME),
@@ -809,7 +754,7 @@ async def _train_nlu_with_validated_data(
             model_type="nlu",
             is_finetuning=model_to_finetune is not None,
         ):
-            await rasa.nlu.train(
+            await rasa.nlu.train.train(
                 config,
                 file_importer,
                 _train_path,
diff --git a/rasa/nlu/__init__.py b/rasa/nlu/__init__.py
index c78df260380d..4defa9baa4c2 100644
--- a/rasa/nlu/__init__.py
+++ b/rasa/nlu/__init__.py
@@ -1,9 +1,6 @@
 import logging
 
 import rasa
-from rasa.nlu.train import train
-from rasa.nlu.test import run_evaluation as test
-from rasa.nlu.test import cross_validate
 
 logging.getLogger(__name__).addHandler(logging.NullHandler())
 
diff --git a/rasa/nlu/test.py b/rasa/nlu/test.py
index 9da46081d065..5bf39964f858 100644
--- a/rasa/nlu/test.py
+++ b/rasa/nlu/test.py
@@ -645,7 +645,7 @@ def _calculate_report(
     report_as_dict: Optional[bool] = None,
     exclude_label: Optional[Text] = None,
 ) -> Tuple[Union[Text, Dict], float, float, float, np.ndarray, List[Text]]:
-    from rasa.test import get_evaluation_metrics
+    from rasa.model_testing import get_evaluation_metrics
     import sklearn.metrics
     import sklearn.utils.multiclass
 
@@ -1843,7 +1843,7 @@ async def compare_nlu(
     Returns: training examples per run
     """
 
-    from rasa.train import train_nlu_async
+    from rasa.model_training import train_nlu_async
 
     training_examples_per_run = []
 
@@ -1930,7 +1930,7 @@ def _compute_metrics(
 
     Returns: metrics
     """
-    from rasa.test import get_evaluation_metrics
+    from rasa.model_testing import get_evaluation_metrics
 
     # compute fold metrics
     targets, predictions = _targets_predictions_from(
@@ -1952,7 +1952,7 @@ def _compute_entity_metrics(
 
     Returns: entity metrics
     """
-    from rasa.test import get_evaluation_metrics
+    from rasa.model_testing import get_evaluation_metrics
 
     entity_metric_results: EntityMetrics = defaultdict(lambda: defaultdict(list))
     extractors = get_entity_extractors(interpreter)
diff --git a/rasa/run.py b/rasa/run.py
deleted file mode 100644
index 2a6c3c8ac55a..000000000000
--- a/rasa/run.py
+++ /dev/null
@@ -1,80 +0,0 @@
-import logging
-import typing
-from typing import Dict, Text
-
-import rasa.shared.utils.common
-from rasa.shared.utils.cli import print_warning
-from rasa.shared.constants import DOCS_BASE_URL
-from rasa.core.lock_store import LockStore
-
-logger = logging.getLogger(__name__)
-
-if typing.TYPE_CHECKING:
-    from rasa.core.agent import Agent
-
-
-def run(
-    model: Text,
-    endpoints: Text,
-    connector: Text = None,
-    credentials: Text = None,
-    **kwargs: Dict,
-):
-    """Runs a Rasa model.
-
-    Args:
-        model: Path to model archive.
-        endpoints: Path to endpoints file.
-        connector: Connector which should be use (overwrites `credentials`
-        field).
-        credentials: Path to channel credentials file.
-        **kwargs: Additional arguments which are passed to
-        `rasa.core.run.serve_application`.
-
-    """
-    import rasa.core.run
-    import rasa.nlu.run
-    from rasa.core.utils import AvailableEndpoints
-
-    _endpoints = AvailableEndpoints.read_endpoints(endpoints)
-
-    if not connector and not credentials:
-        connector = "rest"
-        print_warning(
-            f"No chat connector configured, falling back to the "
-            f"REST input channel. To connect your bot to another channel, "
-            f"read the docs here: {DOCS_BASE_URL}/messaging-and-voice-channels"
-        )
-
-    kwargs = rasa.shared.utils.common.minimal_kwargs(
-        kwargs, rasa.core.run.serve_application
-    )
-    rasa.core.run.serve_application(
-        model,
-        channel=connector,
-        credentials=credentials,
-        endpoints=_endpoints,
-        **kwargs,
-    )
-
-
-def create_agent(model: Text, endpoints: Text = None) -> "Agent":
-    from rasa.core.tracker_store import TrackerStore
-    from rasa.core.utils import AvailableEndpoints
-    from rasa.core.agent import Agent
-    from rasa.core.brokers.broker import EventBroker
-    import rasa.utils.common
-
-    _endpoints = AvailableEndpoints.read_endpoints(endpoints)
-
-    _broker = rasa.utils.common.run_in_loop(EventBroker.create(_endpoints.event_broker))
-    _tracker_store = TrackerStore.create(_endpoints.tracker_store, event_broker=_broker)
-    _lock_store = LockStore.create(_endpoints.lock_store)
-
-    return Agent.load(
-        model,
-        generator=_endpoints.nlg,
-        tracker_store=_tracker_store,
-        lock_store=_lock_store,
-        action_endpoint=_endpoints.action,
-    )
diff --git a/rasa/server.py b/rasa/server.py
index 6e9c08e69dc0..96aac1017768 100644
--- a/rasa/server.py
+++ b/rasa/server.py
@@ -7,7 +7,6 @@
 import traceback
 from collections import defaultdict
 from functools import reduce, wraps
-from http import HTTPStatus
 from inspect import isawaitable
 from pathlib import Path
 from http import HTTPStatus
@@ -68,7 +67,8 @@
 from rasa.shared.core.trackers import DialogueStateTracker, EventVerbosity
 from rasa.core.utils import AvailableEndpoints
 from rasa.nlu.emulators.no_emulator import NoEmulator
-from rasa.nlu.test import run_evaluation, CVEvaluationResult
+import rasa.nlu.test
+from rasa.nlu.test import CVEvaluationResult
 from rasa.utils.endpoints import EndpointConfig
 
 if TYPE_CHECKING:
@@ -1045,7 +1045,7 @@ async def train(request: Request, temporary_directory: Path) -> HTTPResponse:
             with app.active_training_processes.get_lock():
                 app.active_training_processes.value += 1
 
-            from rasa.train import train_async
+            from rasa.model_training import train_async
 
             # pass `None` to run in default executor
             training_result = await train_async(**training_payload)
@@ -1203,7 +1203,7 @@ async def _evaluate_model_using_test_set(
                 HTTPStatus.CONFLICT, "Conflict", "Missing NLU model directory.",
             )
 
-        return await run_evaluation(
+        return await rasa.nlu.test.run_evaluation(
             data_path, nlu_model, disable_plotting=True, report_as_dict=True
         )
 
@@ -1215,7 +1215,7 @@ async def _cross_validate(data_file: Text, config_file: Text, folds: int) -> Dic
         config = await importer.get_config()
         nlu_data = await importer.get_nlu_data()
 
-        evaluations = rasa.nlu.cross_validate(
+        evaluations = rasa.nlu.test.cross_validate(
             data=nlu_data,
             n_folds=folds,
             nlu_config=config,
diff --git a/rasa/shared/core/constants.py b/rasa/shared/core/constants.py
index 41f5ed9d4235..2713b0e1d73e 100644
--- a/rasa/shared/core/constants.py
+++ b/rasa/shared/core/constants.py
@@ -81,4 +81,14 @@
 USE_TEXT_FOR_FEATURIZATION = "use_text_for_featurization"
 ENTITY_LABEL_SEPARATOR = "#"
 
+# if you add more policy/classifier names, make sure to add a test as well to ensure
+# that the name and the class stay in sync
+POLICY_NAME_TWO_STAGE_FALLBACK = "TwoStageFallbackPolicy"
+POLICY_NAME_MAPPING = "MappingPolicy"
+POLICY_NAME_FALLBACK = "FallbackPolicy"
+POLICY_NAME_FORM = "FormPolicy"
+POLICY_NAME_RULE = "RulePolicy"
+
+CLASSIFIER_NAME_FALLBACK = "FallbackClassifier"
+
 POLICIES_THAT_EXTRACT_ENTITIES = {"TEDPolicy"}
diff --git a/rasa/telemetry.py b/rasa/telemetry.py
index e7667bd5f8a0..0883d4b017e7 100644
--- a/rasa/telemetry.py
+++ b/rasa/telemetry.py
@@ -470,11 +470,12 @@ def _default_context_fields() -> Dict[Text, Any]:
     Return:
         A new context containing information about the runtime environment.
     """
-    import tensorflow as tf
 
     global TELEMETRY_CONTEXT
 
     if not TELEMETRY_CONTEXT:
+        # Make sure to update the example in docs/docs/telemetry/telemetry.mdx
+        # if you change / add context
         TELEMETRY_CONTEXT = {
             "os": {"name": platform.system(), "version": platform.release()},
             "ci": in_continuous_integration(),
@@ -482,7 +483,6 @@ def _default_context_fields() -> Dict[Text, Any]:
             "directory": _hash_directory_path(os.getcwd()),
             "python": sys.version.split(" ")[0],
             "rasa_open_source": rasa.__version__,
-            "gpu": len(tf.config.list_physical_devices("GPU")),
             "cpu": multiprocessing.cpu_count(),
             "docker": _is_docker(),
         }
@@ -702,6 +702,8 @@ async def track_model_training(
 
     training_id = uuid.uuid4().hex
 
+    # Make sure to update the example in docs/docs/telemetry/telemetry.mdx
+    # if you change / add any properties
     _track(
         TRAINING_STARTED_EVENT,
         {
diff --git a/rasa/utils/common.py b/rasa/utils/common.py
index 588c8c497615..75354943652b 100644
--- a/rasa/utils/common.py
+++ b/rasa/utils/common.py
@@ -6,7 +6,6 @@
 from types import TracebackType
 from typing import Any, Coroutine, Dict, List, Optional, Text, Type, TypeVar, Union
 
-import rasa.core.utils
 import rasa.utils.io
 from rasa.constants import DEFAULT_LOG_LEVEL_LIBRARIES, ENV_LOG_LEVEL_LIBRARIES
 from rasa.shared.constants import DEFAULT_LOG_LEVEL, ENV_LOG_LEVEL
@@ -104,20 +103,12 @@ def update_tensorflow_log_level() -> None:
     # first import since some warnings are raised on the first import.
     os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
 
-    import tensorflow as tf
-
     log_level = os.environ.get(ENV_LOG_LEVEL_LIBRARIES, DEFAULT_LOG_LEVEL_LIBRARIES)
 
-    if log_level == "DEBUG":
-        tf_log_level = tf.compat.v1.logging.DEBUG
-    elif log_level == "INFO":
-        tf_log_level = tf.compat.v1.logging.INFO
-    elif log_level == "WARNING":
-        tf_log_level = tf.compat.v1.logging.WARN
-    else:
-        tf_log_level = tf.compat.v1.logging.ERROR
+    if not log_level:
+        log_level = "ERROR"
 
-    tf.compat.v1.logging.set_verbosity(tf_log_level)
+    logging.getLogger("tensorflow").setLevel(log_level)
     logging.getLogger("tensorflow").propagate = False
 
 
diff --git a/rasa/utils/train_utils.py b/rasa/utils/train_utils.py
index 81d26b974940..b64b0b5a639f 100644
--- a/rasa/utils/train_utils.py
+++ b/rasa/utils/train_utils.py
@@ -1,5 +1,4 @@
 from pathlib import Path
-import tensorflow as tf
 import copy
 import numpy as np
 from typing import Optional, Text, Dict, Any, Union, List, Tuple, TYPE_CHECKING
@@ -46,6 +45,7 @@
 if TYPE_CHECKING:
     from rasa.nlu.extractors.extractor import EntityTagSpec
     from rasa.nlu.tokenizers.tokenizer import Token
+    from tensorflow.keras.callbacks import Callback
 
 
 def normalize(values: np.ndarray, ranking_length: Optional[int] = 0) -> np.ndarray:
@@ -425,7 +425,7 @@ def create_common_callbacks(
     tensorboard_log_dir: Optional[Text] = None,
     tensorboard_log_level: Optional[Text] = None,
     checkpoint_dir: Optional[Path] = None,
-) -> List[tf.keras.callbacks.Callback]:
+) -> List["Callback"]:
     """Create common callbacks.
 
     The following callbacks are created:
@@ -443,6 +443,8 @@ def create_common_callbacks(
     Returns:
         A list of callbacks.
     """
+    import tensorflow as tf
+
     callbacks = [RasaTrainingLogger(epochs, silent=False)]
 
     if tensorboard_log_dir:
diff --git a/tests/cli/test_cli.py b/tests/cli/test_cli.py
index de6c663051d6..57b3527489cf 100644
--- a/tests/cli/test_cli.py
+++ b/tests/cli/test_cli.py
@@ -1,24 +1,39 @@
+from pathlib import Path
 from typing import Callable
-from _pytest.pytester import RunResult
+from _pytest.pytester import RunResult, Testdir
 import pytest
 import sys
 
 
-def test_cli_start(run: Callable[..., RunResult]):
+def test_cli_start_is_fast(testdir: Testdir):
     """
-    Checks that a call to ``rasa --help`` does not take longer than 7 seconds
-    (10 seconds on Windows).
+    Checks that a call to ``rasa --help`` does not import any slow imports.
+
+    If this is failing this means, that a simple "rasa --help" commands imports
+    `tensorflow` which makes our CLI extremely slow. In case this test is failing
+    you've very likely added a global import of "tensorflow" which should be
+    avoided. Consider making this import (or the import of its parent module)
+    a local import.
+
+    If you are clueless where that import happens, you can run
+    ```
+    python -X importtime -m rasa.__main__ --help  2> import.log
+    tuna import.log
+    ```
+    to get the import chain.
+    (make sure to run with python >= 3.7, and install tune (pip install tuna))
     """
-    import time
 
-    start = time.time()
-    run("--help")
-    end = time.time()
+    rasa_path = str(
+        (Path(__file__).parent / ".." / ".." / "rasa" / "__main__.py").absolute()
+    )
+    args = [sys.executable, "-X", "importtime", rasa_path, "--help"]
+    result = testdir.run(*args)
 
-    duration = end - start
+    assert result.ret == 0
 
-    # it sometimes takes a bit more time to start it on Windows
-    assert duration <= 20 if sys.platform == "win32" else 10
+    # tensorflow is slow -> can't get imported when running basic CLI commands
+    result.stderr.no_fnmatch_line("*tensorflow.python.eager")
 
 
 def test_data_convert_help(run: Callable[..., RunResult]):
diff --git a/tests/cli/test_rasa_interactive.py b/tests/cli/test_rasa_interactive.py
index b49d39c130c0..b87211ae636e 100644
--- a/tests/cli/test_rasa_interactive.py
+++ b/tests/cli/test_rasa_interactive.py
@@ -10,7 +10,7 @@
 from rasa.core.train import do_interactive_learning
 from rasa.core.training import interactive as interactive_learning
 from rasa.cli import interactive, train
-from rasa.train import TrainingResult
+from rasa.model_training import TrainingResult
 from tests.conftest import DEFAULT_NLU_DATA
 
 
diff --git a/tests/cli/test_rasa_train.py b/tests/cli/test_rasa_train.py
index 114439370fac..4fac719fef47 100644
--- a/tests/cli/test_rasa_train.py
+++ b/tests/cli/test_rasa_train.py
@@ -11,7 +11,7 @@
 from rasa import model
 from rasa.nlu.model import Metadata
 from rasa.shared.nlu.training_data import training_data
-from rasa.train import (
+from rasa.model_training import (
     CODE_CORE_NEEDS_TO_BE_RETRAINED,
     CODE_NLU_NEEDS_TO_BE_RETRAINED,
     CODE_NLG_NEEDS_TO_BE_RETRAINED,
diff --git a/tests/conftest.py b/tests/conftest.py
index b30d854caabd..d03216d85f71 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -35,7 +35,7 @@
 import rasa.core.run
 from rasa.core.tracker_store import InMemoryTrackerStore, TrackerStore
 from rasa.model import get_model
-from rasa.train import train_async, train_nlu_async
+from rasa.model_training import train_async, train_nlu_async
 from rasa.utils.common import TempDirectoryPath
 from tests.core.conftest import (
     DEFAULT_DOMAIN_PATH_WITH_SLOTS,
@@ -79,7 +79,7 @@
         Path("tests", "nlu", "selectors").absolute(),
     ],
     "category_full_model_training": [
-        Path("tests", "test_train.py").absolute(),
+        Path("tests", "test_model_training.py").absolute(),
         Path("tests", "nlu", "test_train.py").absolute(),
         Path("tests", "core", "test_training.py").absolute(),
         Path("tests", "core", "test_examples.py").absolute(),
diff --git a/tests/core/policies/test_ted_policy.py b/tests/core/policies/test_ted_policy.py
index 62d5b72f75ed..e0e42c023c21 100644
--- a/tests/core/policies/test_ted_policy.py
+++ b/tests/core/policies/test_ted_policy.py
@@ -22,7 +22,7 @@
 from rasa.utils.tensorflow.data_generator import RasaBatchDataGenerator
 from rasa.shared.core.trackers import DialogueStateTracker
 from rasa.shared.nlu.interpreter import RegexInterpreter
-from rasa.train import train_core
+from rasa.model_training import train_core
 from rasa.utils import train_utils
 from rasa.utils.tensorflow.constants import (
     EVAL_NUM_EXAMPLES,
diff --git a/tests/nlu/classifiers/test_diet_classifier.py b/tests/nlu/classifiers/test_diet_classifier.py
index 47367c748aab..96141340a05c 100644
--- a/tests/nlu/classifiers/test_diet_classifier.py
+++ b/tests/nlu/classifiers/test_diet_classifier.py
@@ -1,5 +1,4 @@
 from pathlib import Path
-from typing import Text
 
 import numpy as np
 import pytest
@@ -9,7 +8,7 @@
 
 import rasa.model
 from rasa.shared.nlu.training_data.features import Features
-from rasa.nlu import train
+import rasa.nlu.train
 from rasa.nlu.classifiers import LABEL_RANKING_LENGTH
 from rasa.nlu.config import RasaNLUModelConfig
 from rasa.shared.nlu.constants import (
@@ -178,7 +177,7 @@ async def _train_persist_load_with_different_settings(
 ):
     _config = RasaNLUModelConfig({"pipeline": pipeline, "language": "en"})
 
-    (trainer, trained, persisted_path) = await train(
+    (trainer, trained, persisted_path) = await rasa.nlu.train.train(
         _config,
         path=str(tmp_path),
         data="data/examples/rasa/demo-rasa-multi-intent.yml",
@@ -287,7 +286,7 @@ async def test_raise_error_on_incorrect_pipeline(component_builder, tmp_path: Pa
     )
 
     with pytest.raises(Exception) as e:
-        await train(
+        await rasa.nlu.train.train(
             _config,
             path=str(tmp_path),
             data=DEFAULT_DATA_PATH,
@@ -351,7 +350,7 @@ async def test_softmax_normalization(
     pipeline[2].update(classifier_params)
 
     _config = RasaNLUModelConfig({"pipeline": pipeline})
-    (trained_model, _, persisted_path) = await train(
+    (trained_model, _, persisted_path) = await rasa.nlu.train.train(
         _config, path=str(tmp_path), data=data_path, component_builder=component_builder
     )
     loaded = Interpreter.load(persisted_path, component_builder)
@@ -399,7 +398,7 @@ async def test_inner_linear_normalization(
     pipeline[2].update(classifier_params)
 
     _config = RasaNLUModelConfig({"pipeline": pipeline})
-    (trained_model, _, persisted_path) = await train(
+    (trained_model, _, persisted_path) = await rasa.nlu.train.train(
         _config, path=str(tmp_path), data=data_path, component_builder=component_builder
     )
     loaded = Interpreter.load(persisted_path, component_builder)
@@ -440,7 +439,7 @@ async def test_margin_loss_is_not_normalized(
     monkeypatch.setattr(train_utils, "normalize", mock.normalize)
 
     _config = RasaNLUModelConfig({"pipeline": pipeline})
-    (trained_model, _, persisted_path) = await train(
+    (trained_model, _, persisted_path) = await rasa.nlu.train.train(
         _config,
         path=str(tmpdir),
         data="data/test/many_intents.yml",
@@ -477,14 +476,14 @@ async def test_set_random_seed(component_builder, tmpdir):
     )
 
     # first run
-    (trained_a, _, persisted_path_a) = await train(
+    (trained_a, _, persisted_path_a) = await rasa.nlu.train.train(
         _config,
         path=tmpdir.strpath + "_a",
         data=DEFAULT_DATA_PATH,
         component_builder=component_builder,
     )
     # second run
-    (trained_b, _, persisted_path_b) = await train(
+    (trained_b, _, persisted_path_b) = await rasa.nlu.train.train(
         _config,
         path=tmpdir.strpath + "_b",
         data=DEFAULT_DATA_PATH,
@@ -525,7 +524,7 @@ async def test_train_tensorboard_logging(
         }
     )
 
-    await train(
+    await rasa.nlu.train.train(
         _config,
         path=str(tmpdir),
         data="data/examples/rasa/demo-rasa-multi-intent.yml",
@@ -562,7 +561,7 @@ async def test_train_model_checkpointing(
         }
     )
 
-    await train(
+    await rasa.nlu.train.train(
         _config,
         path=str(tmpdir),
         data="data/examples/rasa/demo-rasa.yml",
@@ -602,7 +601,7 @@ async def test_train_persist_load_with_composite_entities(
 
     _config = RasaNLUModelConfig({"pipeline": pipeline, "language": "en"})
 
-    (trainer, trained, persisted_path) = await train(
+    (trainer, trained, persisted_path) = await rasa.nlu.train.train(
         _config,
         path=tmpdir.strpath,
         data="data/test/demo-rasa-composite-entities.yml",
diff --git a/tests/nlu/extractors/test_crf_entity_extractor.py b/tests/nlu/extractors/test_crf_entity_extractor.py
index 84849f1046c8..358296c9aab7 100644
--- a/tests/nlu/extractors/test_crf_entity_extractor.py
+++ b/tests/nlu/extractors/test_crf_entity_extractor.py
@@ -5,7 +5,7 @@
 import pytest
 
 from rasa.nlu.components import ComponentBuilder
-from rasa.nlu import train
+import rasa.nlu.train
 from rasa.nlu.config import RasaNLUModelConfig
 from rasa.nlu.model import Interpreter
 from rasa.nlu.featurizers.dense_featurizer.spacy_featurizer import SpacyFeaturizer
@@ -27,7 +27,7 @@ async def test_train_persist_load_with_composite_entities(
 
     _config = RasaNLUModelConfig({"pipeline": pipeline, "language": "en"})
 
-    (trainer, trained, persisted_path) = await train(
+    (trainer, trained, persisted_path) = await rasa.nlu.train.train(
         _config,
         path=str(tmp_path),
         data="data/test/demo-rasa-composite-entities.yml",
@@ -98,7 +98,7 @@ async def test_train_persist_with_different_configurations(
 
     _config = RasaNLUModelConfig({"pipeline": pipeline, "language": "en"})
 
-    (trainer, trained, persisted_path) = await train(
+    (trainer, trained, persisted_path) = await rasa.nlu.train.train(
         _config,
         path=str(tmp_path),
         data="data/examples/rasa",
diff --git a/tests/nlu/selectors/test_selectors.py b/tests/nlu/selectors/test_selectors.py
index 8dbcf1dd0a9b..35c753095334 100644
--- a/tests/nlu/selectors/test_selectors.py
+++ b/tests/nlu/selectors/test_selectors.py
@@ -7,7 +7,7 @@
 from _pytest.monkeypatch import MonkeyPatch
 
 import rasa.model
-from rasa.nlu import train
+import rasa.nlu.train
 from rasa.nlu.components import ComponentBuilder
 from rasa.shared.nlu.training_data import util
 from rasa.nlu.config import RasaNLUModelConfig
@@ -228,7 +228,7 @@ async def test_train_model_checkpointing(
         }
     )
 
-    await train(
+    await rasa.nlu.train.train(
         _config,
         path=str(tmpdir),
         data="data/test_selectors",
@@ -258,7 +258,7 @@ async def _train_persist_load_with_different_settings(
 ):
     _config = RasaNLUModelConfig({"pipeline": pipeline, "language": "en"})
 
-    (trainer, trained, persisted_path) = await train(
+    (trainer, trained, persisted_path) = await rasa.nlu.train.train(
         _config,
         path=str(tmp_path),
         data="data/examples/rasa/demo-rasa.yml",
@@ -342,7 +342,7 @@ async def test_cross_entropy_with_linear_norm(
     pipeline[2].update(classifier_params)
 
     _config = RasaNLUModelConfig({"pipeline": pipeline})
-    (trained_model, _, persisted_path) = await train(
+    (trained_model, _, persisted_path) = await rasa.nlu.train.train(
         _config,
         path=str(tmp_path),
         data="data/test_selectors",
@@ -388,7 +388,7 @@ async def test_margin_loss_is_not_normalized(
     monkeypatch.setattr(train_utils, "normalize", mock.normalize)
 
     _config = RasaNLUModelConfig({"pipeline": pipeline})
-    (trained_model, _, persisted_path) = await train(
+    (trained_model, _, persisted_path) = await rasa.nlu.train.train(
         _config,
         path=str(tmp_path),
         data="data/test_selectors",
@@ -428,7 +428,7 @@ async def test_softmax_ranking(
     pipeline[2].update(classifier_params)
 
     _config = RasaNLUModelConfig({"pipeline": pipeline})
-    (trained_model, _, persisted_path) = await train(
+    (trained_model, _, persisted_path) = await rasa.nlu.train.train(
         _config, path=str(tmp_path), data=data_path, component_builder=component_builder
     )
     loaded = Interpreter.load(persisted_path, component_builder)
diff --git a/tests/nlu/test_components.py b/tests/nlu/test_components.py
index 4c32fb2844ef..e03d0edf4fd3 100644
--- a/tests/nlu/test_components.py
+++ b/tests/nlu/test_components.py
@@ -3,7 +3,8 @@
 
 import pytest
 
-from rasa.nlu import registry, train
+from rasa.nlu import registry
+import rasa.nlu.train
 from rasa.nlu.components import Component, ComponentBuilder, find_unavailable_packages
 from rasa.nlu.config import RasaNLUModelConfig
 from rasa.shared.exceptions import InvalidConfigException
@@ -98,7 +99,7 @@ async def test_example_component(component_builder: ComponentBuilder, tmp_path:
         {"pipeline": [{"name": "tests.nlu.example_component.MyComponent"}]}
     )
 
-    (trainer, trained, persisted_path) = await train(
+    (trainer, trained, persisted_path) = await rasa.nlu.train.train(
         _config,
         data=DEFAULT_DATA_PATH,
         path=str(tmp_path),
@@ -199,7 +200,7 @@ async def test_validate_requirements_raises_exception_on_component_without_name(
     )
 
     with pytest.raises(InvalidConfigException):
-        await train(
+        await rasa.nlu.train.train(
             _config, data=DEFAULT_DATA_PATH, path=str(tmp_path),
         )
 
@@ -211,7 +212,7 @@ async def test_validate_component_keys_raises_warning_on_invalid_key(tmp_path: P
     )
 
     with pytest.warns(UserWarning) as record:
-        await train(
+        await rasa.nlu.train.train(
             _config, data=DEFAULT_DATA_PATH, path=str(tmp_path),
         )
 
diff --git a/tests/nlu/test_evaluation.py b/tests/nlu/test_evaluation.py
index 1b7b42b323bf..e8d665d890d9 100644
--- a/tests/nlu/test_evaluation.py
+++ b/tests/nlu/test_evaluation.py
@@ -12,7 +12,7 @@
 import rasa.shared.nlu.training_data.loading
 import rasa.shared.utils.io
 import rasa.utils.io
-from rasa.nlu import train
+import rasa.nlu.train
 from rasa.nlu.classifiers.diet_classifier import DIETClassifier
 from rasa.nlu.classifiers.fallback_classifier import FallbackClassifier
 from rasa.nlu.components import ComponentBuilder, Component
@@ -65,7 +65,7 @@
 )
 from rasa.shared.nlu.training_data.message import Message
 from rasa.shared.nlu.training_data.training_data import TrainingData
-from rasa.test import compare_nlu_models
+from rasa.model_testing import compare_nlu_models
 from rasa.utils.tensorflow.constants import EPOCHS, ENTITY_RECOGNITION
 from tests.nlu.conftest import DEFAULT_DATA_PATH
 
@@ -394,7 +394,7 @@ async def test_eval_data(
         ],
     )
 
-    (_, _, persisted_path) = await train(
+    (_, _, persisted_path) = await rasa.nlu.train.train(
         _config,
         path=str(tmp_path),
         data=data_importer,
diff --git a/tests/nlu/test_persistor.py b/tests/nlu/test_persistor.py
index baa0a2ea18d0..d7d086168abf 100644
--- a/tests/nlu/test_persistor.py
+++ b/tests/nlu/test_persistor.py
@@ -4,7 +4,8 @@
 
 from moto import mock_s3
 
-from rasa.nlu import persistor, train
+from rasa.nlu import persistor
+import rasa.nlu.train
 from rasa.nlu.config import RasaNLUModelConfig
 
 
@@ -24,7 +25,7 @@ async def test_list_method_method_in_AWS_persistor(  # noinspection PyPep8Naming
         os.environ["BUCKET_NAME"] = "rasa-test"
         os.environ["AWS_DEFAULT_REGION"] = "us-west-1"
 
-        (trained, _, persisted_path) = await train(
+        (trained, _, persisted_path) = await rasa.nlu.train.train(
             _config,
             data="data/test/demo-rasa-small.json",
             path=str(tmp_path),
diff --git a/tests/nlu/test_train.py b/tests/nlu/test_train.py
index 93924ede2f9d..2fdefc196991 100644
--- a/tests/nlu/test_train.py
+++ b/tests/nlu/test_train.py
@@ -1,7 +1,8 @@
 import os
 import pytest
 
-from rasa.nlu import registry, train
+from rasa.nlu import registry
+import rasa.nlu.train
 from rasa.nlu.config import RasaNLUModelConfig
 from rasa.nlu.model import Interpreter, Trainer
 from rasa.shared.nlu.training_data.training_data import TrainingData
@@ -138,7 +139,7 @@ def test_all_components_are_in_at_least_one_test_pipeline():
 async def test_train_persist_load_parse(language, pipeline, component_builder, tmpdir):
     _config = RasaNLUModelConfig({"pipeline": pipeline, "language": language})
 
-    (trained, _, persisted_path) = await train(
+    (trained, _, persisted_path) = await rasa.nlu.train.train(
         _config,
         path=tmpdir.strpath,
         data=DEFAULT_DATA_PATH,
@@ -212,7 +213,7 @@ async def test_train_model_empty_pipeline(component_builder):
     _config = RasaNLUModelConfig({"pipeline": None, "language": "en"})
 
     with pytest.raises(ValueError):
-        await train(
+        await rasa.nlu.train.train(
             _config, data=DEFAULT_DATA_PATH, component_builder=component_builder
         )
 
@@ -222,7 +223,7 @@ async def test_train_named_model(component_builder, tmpdir):
         {"pipeline": [{"name": "KeywordIntentClassifier"}], "language": "en"}
     )
 
-    (trained, _, persisted_path) = await train(
+    (trained, _, persisted_path) = await rasa.nlu.train.train(
         _config,
         path=tmpdir.strpath,
         data=DEFAULT_DATA_PATH,
@@ -242,7 +243,7 @@ async def test_handles_pipeline_with_non_existing_component(
     pretrained_embeddings_spacy_config.pipeline.append({"name": "my_made_up_component"})
 
     with pytest.raises(Exception) as execinfo:
-        await train(
+        await rasa.nlu.train.train(
             pretrained_embeddings_spacy_config,
             data=DEFAULT_DATA_PATH,
             component_builder=component_builder,
@@ -255,7 +256,7 @@ async def test_train_model_training_data_persisted(component_builder, tmpdir):
         {"pipeline": [{"name": "KeywordIntentClassifier"}], "language": "en"}
     )
 
-    (trained, _, persisted_path) = await train(
+    (trained, _, persisted_path) = await rasa.nlu.train.train(
         _config,
         path=tmpdir.strpath,
         data=DEFAULT_DATA_PATH,
@@ -276,7 +277,7 @@ async def test_train_model_no_training_data_persisted(component_builder, tmpdir)
         {"pipeline": [{"name": "KeywordIntentClassifier"}], "language": "en"}
     )
 
-    (trained, _, persisted_path) = await train(
+    (trained, _, persisted_path) = await rasa.nlu.train.train(
         _config,
         path=tmpdir.strpath,
         data=DEFAULT_DATA_PATH,
diff --git a/tests/shared/core/test_constants.py b/tests/shared/core/test_constants.py
new file mode 100644
index 000000000000..1906d7a7c908
--- /dev/null
+++ b/tests/shared/core/test_constants.py
@@ -0,0 +1,41 @@
+from typing import Text, Type
+
+import pytest
+
+from rasa.core.policies.fallback import FallbackPolicy
+from rasa.core.policies.form_policy import FormPolicy
+from rasa.core.policies.mapping_policy import MappingPolicy
+from rasa.core.policies.rule_policy import RulePolicy
+from rasa.core.policies.two_stage_fallback import TwoStageFallbackPolicy
+from rasa.nlu.classifiers.fallback_classifier import FallbackClassifier
+from rasa.shared.core.constants import (
+    CLASSIFIER_NAME_FALLBACK,
+    POLICY_NAME_FALLBACK,
+    POLICY_NAME_MAPPING,
+    POLICY_NAME_RULE,
+    POLICY_NAME_TWO_STAGE_FALLBACK,
+    POLICY_NAME_FORM,
+)
+
+
+@pytest.mark.parametrize(
+    "name_in_constant, policy_class",
+    [
+        (POLICY_NAME_TWO_STAGE_FALLBACK, TwoStageFallbackPolicy),
+        (POLICY_NAME_FALLBACK, FallbackPolicy),
+        (POLICY_NAME_MAPPING, MappingPolicy),
+        (POLICY_NAME_FORM, FormPolicy),
+        (POLICY_NAME_RULE, RulePolicy),
+        (CLASSIFIER_NAME_FALLBACK, FallbackClassifier),
+    ],
+)
+def test_policy_names(name_in_constant: Text, policy_class: Type):
+    assert name_in_constant == policy_class.__name__
+
+
+@pytest.mark.parametrize(
+    "name_in_constant, classifier_class",
+    [(CLASSIFIER_NAME_FALLBACK, FallbackClassifier),],
+)
+def test_classifier_names(name_in_constant: Text, classifier_class: Type):
+    assert name_in_constant == classifier_class.__name__
diff --git a/tests/test_model.py b/tests/test_model.py
index 52e97902cf95..25ba91aad9ee 100644
--- a/tests/test_model.py
+++ b/tests/test_model.py
@@ -51,7 +51,7 @@
     FingerprintComparisonResult,
 )
 from rasa.exceptions import ModelNotFound
-from rasa.train import train_core, train_core_async
+from rasa.model_training import train_core, train_core_async
 from tests.core.conftest import DEFAULT_DOMAIN_PATH_WITH_MAPPING, DEFAULT_STACK_CONFIG
 
 
diff --git a/tests/test_test.py b/tests/test_model_testing.py
similarity index 97%
rename from tests/test_test.py
rename to tests/test_model_testing.py
index aea5aed8b157..a753218186ed 100644
--- a/tests/test_test.py
+++ b/tests/test_model_testing.py
@@ -35,7 +35,7 @@ def monkeypatch_get_latest_model(tmp_path: Path, monkeypatch: MonkeyPatch) -> No
 def test_get_sanitized_model_directory_when_not_passing_model(
     capsys: CaptureFixture, tmp_path: Path, monkeypatch: MonkeyPatch
 ):
-    from rasa.test import _get_sanitized_model_directory
+    from rasa.model_testing import _get_sanitized_model_directory
 
     monkeypatch_get_latest_model(tmp_path, monkeypatch)
 
@@ -54,7 +54,7 @@ def test_get_sanitized_model_directory_when_not_passing_model(
 def test_get_sanitized_model_directory_when_passing_model_file_explicitly(
     capsys: CaptureFixture, tmp_path: Path, monkeypatch: MonkeyPatch
 ):
-    from rasa.test import _get_sanitized_model_directory
+    from rasa.model_testing import _get_sanitized_model_directory
 
     monkeypatch_get_latest_model(tmp_path, monkeypatch)
 
@@ -73,7 +73,7 @@ def test_get_sanitized_model_directory_when_passing_model_file_explicitly(
 def test_get_sanitized_model_directory_when_passing_other_input(
     capsys: CaptureFixture, tmp_path: Path, monkeypatch: MonkeyPatch
 ):
-    from rasa.test import _get_sanitized_model_directory
+    from rasa.model_testing import _get_sanitized_model_directory
 
     monkeypatch_get_latest_model(tmp_path, monkeypatch)
 
@@ -109,7 +109,7 @@ def test_get_sanitized_model_directory_when_passing_other_input(
 def test_get_evaluation_metrics(
     targets, predictions, expected_precision, expected_fscore, expected_accuracy
 ):
-    from rasa.test import get_evaluation_metrics
+    from rasa.model_testing import get_evaluation_metrics
 
     report, precision, f1, accuracy = get_evaluation_metrics(
         targets, predictions, True, exclude_label=NO_ENTITY
@@ -140,7 +140,7 @@ def test_get_evaluation_metrics(
     ],
 )
 def test_get_label_set(targets, exclude_label, expected):
-    from rasa.test import get_unique_labels
+    from rasa.model_testing import get_unique_labels
 
     actual = get_unique_labels(targets, exclude_label)
     assert set(expected) == set(actual)
@@ -158,7 +158,7 @@ async def test_interpreter_passed_to_agent(
 def test_e2e_warning_if_no_nlu_model(
     monkeypatch: MonkeyPatch, trained_core_model: Text, capsys: CaptureFixture
 ):
-    from rasa.test import test_core
+    from rasa.model_testing import test_core
 
     # Patching is bit more complicated as we have a module `train` and function
     # with the same name 😬
diff --git a/tests/test_train.py b/tests/test_model_training.py
similarity index 92%
rename from tests/test_train.py
rename to tests/test_model_training.py
index 2d5145d9d8e6..0e0354e838d1 100644
--- a/tests/test_train.py
+++ b/tests/test_model_training.py
@@ -15,6 +15,7 @@
 from rasa.core.policies.ted_policy import TEDPolicy
 import rasa.model
 import rasa.core
+import rasa.core.train
 import rasa.nlu
 from rasa.nlu.classifiers.diet_classifier import DIETClassifier
 import rasa.shared.importers.autoconfig as autoconfig
@@ -23,7 +24,7 @@
 from rasa.core.interpreter import RasaNLUInterpreter
 from rasa.nlu.model import Interpreter
 
-from rasa.train import train_core, train_nlu, train, dry_run_result
+from rasa.model_training import train_core, train_nlu, train, dry_run_result
 from rasa.utils.tensorflow.constants import EPOCHS
 from tests.conftest import DEFAULT_CONFIG_PATH, DEFAULT_NLU_DATA, AsyncMock
 from tests.core.conftest import DEFAULT_DOMAIN_PATH_WITH_SLOTS, DEFAULT_STORIES_FILE
@@ -233,14 +234,13 @@ def test_trained_interpreter_passed_to_core_training(
     # Patching is bit more complicated as we have a module `train` and function
     # with the same name 😬
     monkeypatch.setattr(
-        sys.modules["rasa.train"],
+        rasa.model_training,
         "_train_nlu_with_validated_data",
         AsyncMock(return_value=unpacked_trained_rasa_model),
     )
 
     # Mock the actual Core training
-    _train_core = AsyncMock()
-    monkeypatch.setattr(rasa.core, "train", _train_core)
+    _train_core = mock_core_training(monkeypatch)
 
     train(
         DEFAULT_DOMAIN_PATH_WITH_SLOTS,
@@ -270,8 +270,7 @@ def test_interpreter_of_old_model_passed_to_core_training(
     )
 
     # Mock the actual Core training
-    _train_core = AsyncMock()
-    monkeypatch.setattr(rasa.core, "train", _train_core)
+    _train_core = mock_core_training(monkeypatch)
 
     train(
         DEFAULT_DOMAIN_PATH_WITH_SLOTS,
@@ -286,13 +285,13 @@ def test_interpreter_of_old_model_passed_to_core_training(
 
 
 def test_load_interpreter_returns_none_for_none():
-    from rasa.train import _load_interpreter
+    from rasa.model_training import _load_interpreter
 
     assert _load_interpreter(None) is None
 
 
 def test_interpreter_from_previous_model_returns_none_for_none():
-    from rasa.train import _interpreter_from_previous_model
+    from rasa.model_training import _interpreter_from_previous_model
 
     assert _interpreter_from_previous_model(None) is None
 
@@ -312,7 +311,7 @@ def test_train_core_autoconfig(
 
     # skip actual core training
     monkeypatch.setattr(
-        sys.modules["rasa.train"], "_train_core_with_validated_data", AsyncMock()
+        rasa.model_training, "_train_core_with_validated_data", AsyncMock()
     )
 
     # do training
@@ -341,7 +340,7 @@ def test_train_nlu_autoconfig(
     monkeypatch.setattr(autoconfig, "get_configuration", mocked_get_configuration)
 
     monkeypatch.setattr(
-        sys.modules["rasa.train"], "_train_nlu_with_validated_data", AsyncMock()
+        rasa.model_training, "_train_nlu_with_validated_data", AsyncMock()
     )
 
     # do training
@@ -367,11 +366,11 @@ async def mock_async_func(*args: Any, **kwargs: Any) -> None:
 
 
 def mock_core_training(monkeypatch: MonkeyPatch) -> Mock:
-    return mock_async(monkeypatch, rasa.core, rasa.core.train.__name__)
+    return mock_async(monkeypatch, rasa.core.train, rasa.core.train.train.__name__)
 
 
 def mock_nlu_training(monkeypatch: MonkeyPatch) -> Mock:
-    return mock_async(monkeypatch, rasa.nlu, rasa.nlu.train.__name__)
+    return mock_async(monkeypatch, rasa.nlu.train, rasa.nlu.train.train.__name__)
 
 
 def new_model_path_in_same_dir(old_model_path: Text) -> Text:
@@ -620,11 +619,8 @@ def test_model_finetuning(
     trained_rasa_model: Text,
     use_latest_model: bool,
 ):
-    mocked_nlu_training = Mock(wraps=rasa.nlu.train)
-    monkeypatch.setattr(rasa.nlu, rasa.nlu.train.__name__, mocked_nlu_training)
-
-    mocked_core_training = Mock(wraps=rasa.core.train)
-    monkeypatch.setattr(rasa.core, rasa.core.train.__name__, mocked_core_training)
+    mocked_nlu_training = mock_nlu_training(monkeypatch)
+    mocked_core_training = mock_core_training(monkeypatch)
 
     (tmp_path / "models").mkdir()
     output = str(tmp_path / "models")
@@ -659,9 +655,7 @@ def test_model_finetuning_core(
     trained_moodbot_path: Text,
     use_latest_model: bool,
 ):
-    mocked_core_training = AsyncMock()
-    monkeypatch.setattr(rasa.core, rasa.core.train.__name__, mocked_core_training)
-
+    mocked_core_training = mock_core_training(monkeypatch)
     mock_agent_load = Mock(wraps=Agent.load)
     monkeypatch.setattr(Agent, "load", mock_agent_load)
 
@@ -710,9 +704,7 @@ def test_model_finetuning_core(
 def test_model_finetuning_core_with_default_epochs(
     tmp_path: Path, monkeypatch: MonkeyPatch, trained_moodbot_path: Text,
 ):
-    mocked_core_training = AsyncMock()
-    monkeypatch.setattr(rasa.core, rasa.core.train.__name__, mocked_core_training)
-
+    mocked_core_training = mock_core_training(monkeypatch)
     (tmp_path / "models").mkdir()
     output = str(tmp_path / "models")
 
@@ -743,8 +735,7 @@ def test_model_finetuning_core_with_default_epochs(
 def test_model_finetuning_core_new_domain_label(
     tmp_path: Path, monkeypatch: MonkeyPatch, trained_moodbot_path: Text,
 ):
-    mocked_core_training = AsyncMock()
-    monkeypatch.setattr(rasa.core, rasa.core.train.__name__, mocked_core_training)
+    mocked_core_training = mock_core_training(monkeypatch)
 
     (tmp_path / "models").mkdir()
     output = str(tmp_path / "models")
@@ -770,10 +761,8 @@ def test_model_finetuning_core_new_domain_label(
 def test_model_finetuning_new_domain_label_stops_all_training(
     tmp_path: Path, monkeypatch: MonkeyPatch, trained_moodbot_path: Text,
 ):
-    mocked_core_training = AsyncMock()
-    mocked_nlu_training = AsyncMock()
-    monkeypatch.setattr(rasa.core, rasa.core.train.__name__, mocked_core_training)
-    monkeypatch.setattr(rasa.nlu, rasa.nlu.train.__name__, mocked_nlu_training)
+    mocked_core_training = mock_core_training(monkeypatch)
+    mocked_nlu_training = mock_nlu_training(monkeypatch)
 
     (tmp_path / "models").mkdir()
     output = str(tmp_path / "models")
@@ -807,8 +796,7 @@ def test_model_finetuning_nlu(
     trained_nlu_moodbot_path: Text,
     use_latest_model: bool,
 ):
-    mocked_nlu_training = AsyncMock(return_value="")
-    monkeypatch.setattr(rasa.nlu, rasa.nlu.train.__name__, mocked_nlu_training)
+    mocked_nlu_training = mock_nlu_training(monkeypatch)
 
     mock_interpreter_create = Mock(wraps=Interpreter.create)
     monkeypatch.setattr(Interpreter, "create", mock_interpreter_create)
@@ -862,8 +850,7 @@ def test_model_finetuning_nlu(
 def test_model_finetuning_nlu_new_label(
     tmp_path: Path, monkeypatch: MonkeyPatch, trained_nlu_moodbot_path: Text,
 ):
-    mocked_nlu_training = AsyncMock(return_value="")
-    monkeypatch.setattr(rasa.nlu, rasa.nlu.train.__name__, mocked_nlu_training)
+    mocked_nlu_training = mock_nlu_training(monkeypatch)
 
     (tmp_path / "models").mkdir()
     output = str(tmp_path / "models")
@@ -888,8 +875,7 @@ def test_model_finetuning_nlu_new_label(
 def test_model_finetuning_nlu_new_entity(
     tmp_path: Path, monkeypatch: MonkeyPatch, trained_nlu_moodbot_path: Text,
 ):
-    mocked_nlu_training = AsyncMock(return_value="")
-    monkeypatch.setattr(rasa.nlu, rasa.nlu.train.__name__, mocked_nlu_training)
+    mocked_nlu_training = mock_nlu_training(monkeypatch)
 
     (tmp_path / "models").mkdir()
     output = str(tmp_path / "models")
@@ -919,8 +905,7 @@ def test_model_finetuning_nlu_new_label_already_in_domain(
     default_config_path: Text,
     default_domain_path: Text,
 ):
-    mocked_nlu_training = AsyncMock(return_value="")
-    monkeypatch.setattr(rasa.nlu, rasa.nlu.train.__name__, mocked_nlu_training)
+    mocked_nlu_training = mock_nlu_training(monkeypatch)
 
     (tmp_path / "models").mkdir()
     output = str(tmp_path / "models")
@@ -946,8 +931,7 @@ def test_model_finetuning_nlu_new_label_already_in_domain(
 def test_model_finetuning_nlu_new_label_to_domain_only(
     tmp_path: Path, monkeypatch: MonkeyPatch, trained_nlu_moodbot_path: Text,
 ):
-    mocked_nlu_training = AsyncMock(return_value="")
-    monkeypatch.setattr(rasa.nlu, rasa.nlu.train.__name__, mocked_nlu_training)
+    mocked_nlu_training = mock_nlu_training(monkeypatch)
 
     (tmp_path / "models").mkdir()
     output = str(tmp_path / "models")
@@ -972,8 +956,7 @@ def test_model_finetuning_nlu_new_label_to_domain_only(
 def test_model_finetuning_nlu_with_default_epochs(
     tmp_path: Path, monkeypatch: MonkeyPatch, trained_nlu_moodbot_path: Text,
 ):
-    mocked_nlu_training = AsyncMock(return_value="")
-    monkeypatch.setattr(rasa.nlu, rasa.nlu.train.__name__, mocked_nlu_training)
+    mocked_nlu_training = mock_nlu_training(monkeypatch)
 
     (tmp_path / "models").mkdir()
     output = str(tmp_path / "models")
@@ -1012,12 +995,9 @@ def test_model_finetuning_with_invalid_model(
     model_to_fine_tune: Text,
     capsys: CaptureFixture,
 ):
-    mocked_nlu_training = AsyncMock(return_value="")
-    monkeypatch.setattr(rasa.nlu, rasa.nlu.train.__name__, mocked_nlu_training)
-
-    mocked_core_training = AsyncMock()
-    monkeypatch.setattr(rasa.core, rasa.core.train.__name__, mocked_core_training)
+    mocked_nlu_training = mock_nlu_training(monkeypatch)
 
+    mocked_core_training = mock_core_training(monkeypatch)
     (tmp_path / "models").mkdir()
     output = str(tmp_path / "models")
 
@@ -1048,9 +1028,7 @@ def test_model_finetuning_with_invalid_model_core(
     model_to_fine_tune: Text,
     capsys: CaptureFixture,
 ):
-    mocked_core_training = AsyncMock()
-    monkeypatch.setattr(rasa.core, rasa.core.train.__name__, mocked_core_training)
-
+    mocked_core_training = mock_core_training(monkeypatch)
     (tmp_path / "models").mkdir()
     output = str(tmp_path / "models")
 
@@ -1079,8 +1057,7 @@ def test_model_finetuning_with_invalid_model_nlu(
     model_to_fine_tune: Text,
     capsys: CaptureFixture,
 ):
-    mocked_nlu_training = AsyncMock(return_value="")
-    monkeypatch.setattr(rasa.nlu, rasa.nlu.train.__name__, mocked_nlu_training)
+    mocked_nlu_training = mock_nlu_training(monkeypatch)
 
     (tmp_path / "models").mkdir()
     output = str(tmp_path / "models")
diff --git a/tests/test_server.py b/tests/test_server.py
index c6270b7c372a..5df96ca2218d 100644
--- a/tests/test_server.py
+++ b/tests/test_server.py
@@ -43,6 +43,7 @@
 from rasa.core.channels.slack import SlackBot
 from rasa.core.tracker_store import InMemoryTrackerStore
 from rasa.model import unpack_model
+import rasa.nlu.test
 from rasa.nlu.test import CVEvaluationResult
 from rasa.shared.core import events
 from rasa.shared.core.constants import (
@@ -62,7 +63,7 @@
 )
 from rasa.shared.core.trackers import DialogueStateTracker
 from rasa.shared.nlu.constants import INTENT_NAME_KEY
-from rasa.train import TrainingResult
+from rasa.model_training import TrainingResult
 from rasa.utils.endpoints import EndpointConfig
 from tests.core.conftest import DEFAULT_STACK_CONFIG
 from tests.nlu.utilities import ResponseTest
@@ -217,7 +218,7 @@ def run_server(monkeypatch: MonkeyPatch) -> NoReturn:
         import sys
 
         monkeypatch.setattr(
-            sys.modules["rasa.train"], "train_async", mocked_training_function,
+            sys.modules["rasa.model_training"], "train_async", mocked_training_function,
         )
 
         from rasa import __main__
@@ -990,7 +991,9 @@ async def test_cross_validation_with_callback_success(
             )
         )
         monkeypatch.setattr(
-            rasa.nlu, rasa.nlu.cross_validate.__name__, mocked_cross_validation
+            rasa.nlu.test,
+            rasa.nlu.test.cross_validate.__name__,
+            mocked_cross_validation,
         )
 
         _, response = await rasa_app_nlu.post(
@@ -1033,7 +1036,9 @@ async def test_cross_validation_with_callback_error(
     payload = f"{nlu_data}\n{config}"
 
     monkeypatch.setattr(
-        rasa.nlu, rasa.nlu.cross_validate.__name__, Mock(side_effect=ValueError())
+        rasa.nlu.test,
+        rasa.nlu.test.cross_validate.__name__,
+        Mock(side_effect=ValueError()),
     )
 
     callback_url = "https://example.com/webhooks/actions"