Skip to main content

How to Save a scikit-learn Model in the skops Format with MLflow

· 6 min read
Jagdish Kumawat
Founder @ Dewiride

MLflow can save a scikit-learn model in three formats: pickle, cloudpickle and skops. Only skops checks what it builds when the model is loaded, which is why scikit-learn and MLflow recommend it. For most models it needs nothing from you; for models that hold a type skops doesn't trust by default, you list that type once, when logging. This guide logs such a model, with real output from MLflow 3.18.0.

Syntax​

Syntax
mlflow.sklearn.log_model(
sk_model,
name="...",
serialization_format="skops",
skops_trusted_types=["module.path.TypeName", ...],
)
PartWhat it means
serialization_format="skops"Save the model as model.skops. Outside Databricks it's already the default; writing it out makes the choice visible
skops_trusted_typesThe types in the model that skops doesn't trust by default and that you reviewed. MLflow saves the list in MLmodel and loads the model with it

MLflow's reference on the formats: "The 'skops' format guarantees safe deserialization", while cloudpickle "requires exercising caution because these formats rely on Python's object serialization mechanism, which can execute arbitrary code during deserialization."

Prerequisites​

  • An MLflow tracking server at http://127.0.0.1:8080. MLflow's tracking server guide shows how to start one.
  • mlflow and scikit-learn in your project, for example with uv add mlflow scikit-learn pandas. skops is installed with MLflow.

Step 1: Log a Model That Needs a Trusted Type​

A HistGradientBoostingRegressor keeps each of its trees in a TreePredictor, a type skops doesn't trust by default; How to Find the Untrusted Types in a scikit-learn Model with skops shows how to find such types. The example trains one on scikit-learn's diabetes data, logs it in the skops format with that type listed, then loads it back:

The skops_model.py script in a macOS-style code window: a TRUSTED_TYPES list holding TreePredictor, a HistGradientBoostingRegressor fitted to the diabetes data, log_model with an input example, serialization_format skops and skops_trusted_types, then mlflow.sklearn.load_model and three predictions.

skops_model.py
import mlflow
from sklearn.datasets import load_diabetes
from sklearn.ensemble import HistGradientBoostingRegressor

mlflow.set_tracking_uri("http://127.0.0.1:8080")
mlflow.set_experiment("skops-demo")

# The one type in this model that skops doesn't trust by default, reviewed and allowed.
TRUSTED_TYPES = ["sklearn.ensemble._hist_gradient_boosting.predictor.TreePredictor"]

X, y = load_diabetes(return_X_y=True, as_frame=True)
model = HistGradientBoostingRegressor(random_state=0).fit(X, y)

with mlflow.start_run():
model_info = mlflow.sklearn.log_model(
model,
name="diabetes-model",
input_example=X.head(3),
serialization_format="skops",
skops_trusted_types=TRUSTED_TYPES,
)

loaded = mlflow.sklearn.load_model(model_info.model_uri)
print(type(loaded).__name__, loaded.predict(X.head(3)).round(1))
Terminal
uv run python skops_model.py
Output
2026/11/21 14:12:18 INFO mlflow.tracking.fluent: Experiment with name 'skops-demo' does not exist. Creating a new experiment.
2026/11/21 14:12:19 INFO mlflow.utils.uv_utils: Detected uv project: found uv.lock and pyproject.toml in /Users/you/source/mlflow-demo
2026/11/21 14:12:19 INFO mlflow.utils.uv_utils: Detected uv project: found uv.lock and pyproject.toml in /Users/you/source/mlflow-demo
2026/11/21 14:12:19 INFO mlflow.utils.environment: Detected uv project at /Users/you/source/mlflow-demo. Attempting to export requirements via 'uv export'.
2026/11/21 14:12:19 INFO mlflow.utils.uv_utils: Exported 91 dependencies via uv
2026/11/21 14:12:19 INFO mlflow.utils.environment: Successfully exported 91 requirements from uv project. Skipping package capture based inference.
🏃 View run casual-mole-223 at: http://127.0.0.1:8080/#/experiments/8/runs/524464d390184c2791bc6acbd2a1e72d
🧪 View experiment at: http://127.0.0.1:8080/#/experiments/8
HistGradientBoostingRegressor [167.3 75.9 133.3]

The output of skops_model.py in a terminal: the requirements export and the run's link, then the loaded HistGradientBoostingRegressor predicts 167.3, 75.9 and 133.3.

The model was saved and loaded back without any warning, and load_model didn't need the trusted list: MLflow read it from the model.

Step 2: See the Format in MLmodel​

Open the model from the experiment's Models tab, then Artifacts and MLmodel:

The model's Artifacts tab with MLmodel open: the file list shows model.skops, and under the sklearn flavor pickled_model is model.skops, serialization_format skops, and skops_trusted_types lists sklearn.ensemble._hist_gradient_boosting.predictor.TreePredictor.

MLmodel (excerpt)
sklearn:
code: null
pickled_model: model.skops
serialization_format: skops
sklearn_version: 1.9.1
skops_trusted_types:
- sklearn.ensemble._hist_gradient_boosting.predictor.TreePredictor

That list is also why skops isn't a free pass: whoever logs a model writes its trusted types, and anyone loading it gets that list. Load models only from people you trust, and see How to Refuse to Load Untrusted Models in MLflow for the setting that blocks loading altogether.

Common Mistakes​

MistakeWhat happensDo this instead
Logging a HistGradientBoosting* model without skops_trusted_typesMlflowException: The saved sklearn model references untrusted typesList TreePredictor after reviewing it
Trusting everything get_untrusted_types reports, unreadThe check no longer protects anyoneReview each type and write out the list
Turning on skops through autologging for such a modelAutologging has no trusted types setting, and the model isn't savedFixed: Encountered unexpected error during sklearn autologging

Troubleshooting​

SymptomLikely causeFix
MlflowException: The saved sklearn model references untrusted types...A type skops doesn't trust by defaultFixed: The saved sklearn model references untrusted types
WARNING mlflow.sklearn: Saving scikit-learn models in the pickle or cloudpickle format requires exercising caution...The model is saved as cloudpickleFixed: the pickle or cloudpickle caution

FAQ​

Do I need skops_trusted_types for every model?​

No. A linear regression, for example, holds only types skops trusts: How to Log a scikit-learn Model in MLflow logged one in the skops format with skops_trusted_types: null.

Do I pass the trusted types again when loading?​

No. MLflow saved them in MLmodel and used them: mlflow.sklearn.load_model above took only the model URI.

Conclusion​

serialization_format="skops" saved the gradient-boosted model as model.skops, and skops_trusted_types let skops build its TreePredictor objects. MLflow kept the list in MLmodel, so loading the model needed nothing more than its URI.

Additional Resources​