mlflow/mlflow · error · TypeError

outputs must be either None, mlflow.models.signature.Schema,

Error message

outputs must be either None, mlflow.models.signature.Schema, or a dataclass,got '{type(outputs).__name__}'

What it means

A TypeError raised by `ModelSignature.__init__` in mlflow/models/signature.py:87 when the `outputs` argument is neither None, a `Schema`, nor a dataclass. It mirrors the inputs validation: only explicitly supported types are accepted so signature metadata can be reliably serialized.

Source

Thrown at mlflow/models/signature.py:87

    dataset, model predictions using and params for inference, or constructed by hand by
    passing an input and output :py:class:`Schema <mlflow.types.Schema>`, and params
    :py:class:`ParamSchema <mlflow.types.ParamSchema>`.
    """

    def __init__(
        self,
        # `dataclass` is an invalid type annotation. Use `Any` instead as a workaround.
        inputs: Schema | Any = None,
        outputs: Schema | Any = None,
        params: ParamSchema = None,
    ):
        if inputs and not isinstance(inputs, Schema) and not is_dataclass(inputs):
            raise TypeError(
                "inputs must be either None, mlflow.models.signature.Schema, or a dataclass,"
                f"got '{type(inputs).__name__}'"
            )
        if outputs and not isinstance(outputs, Schema) and not is_dataclass(outputs):
            raise TypeError(
                "outputs must be either None, mlflow.models.signature.Schema, or a dataclass,"
                f"got '{type(outputs).__name__}'"
            )
        if params and not isinstance(params, ParamSchema):
            raise TypeError(
                "If params are provided, they must by of type mlflow.models.signature.ParamSchema, "
                f"got '{type(params).__name__}'"
            )
        if all(x is None for x in [inputs, outputs, params]):
            raise ValueError("At least one of inputs, outputs or params must be provided")
        if is_dataclass(inputs):
            self.inputs = convert_dataclass_to_schema(inputs)
        else:
            self.inputs = inputs
        if is_dataclass(outputs):
            self.outputs = convert_dataclass_to_schema(outputs)
        else:
            self.outputs = outputs

View on GitHub (pinned to 6a27f2decc)

Solutions

Build the signature with `infer_signature(model_input, model_output)` so outputs are converted to a Schema automatically.
Pass an explicit `mlflow.models.signature.Schema` for outputs.
Pass a dataclass instance describing outputs if you prefer typed definitions.

Example fix

// before
ModelSignature(inputs=schema, outputs=y_array)  # ndarray not allowed
// after
sig = infer_signature(X, y_array)

Defensive patterns

Strategy: type-guard

Validate before calling

from mlflow.models.signature import Schema
from dataclasses import is_dataclass

def validate_outputs_arg(outputs) -> None:
    if outputs and not isinstance(outputs, Schema) and not is_dataclass(outputs):
        raise TypeError(f'outputs must be Schema or dataclass, got {type(outputs).__name__}')

Type guard

def is_valid_signature_outputs(x) -> bool:
    from mlflow.models.signature import Schema
    from dataclasses import is_dataclass
    return x is None or isinstance(x, Schema) or is_dataclass(x)

Try / catch

try:
    sig = ModelSignature(inputs=schema, outputs=raw_predictions)
except TypeError:
    sig = infer_signature(X, raw_predictions)

Prevention

Derive outputs from infer_signature(model_input, model_output).
Never pass raw arrays/DataFrames as outputs to ModelSignature.
Add a constructor-level assertion in helper code that wraps signature building.

When it happens

Trigger: Constructing `ModelSignature(inputs=..., outputs=...)` with raw example outputs (DataFrame, array, dict, list) rather than a Schema or dataclass instance.

Common situations: Manually assembling signatures with prediction output samples, refactored code that swapped infer_signature for direct construction.

Understand the failure class

Background: Invalid argument type errors: "must be of type string", "expected X, got Y", and ERR_INVALID_ARG_TYPE explained — this error's family across 15 libraries.

Related errors

AI-assisted analysis of mlflow/mlflow@6a27f2decc (2026-08-29). Data as JSON: /api/errors/b8aeb4bf951ff66a. Report an issue: GitHub.