Missing blank line between this import and the # Module-level logger comment. PEP 8 and the existing file style expect a blank line separating import groups from module-level code.

from sagemaker.train.common_utils.finetune_utils import _resolve_mlflow_resource_arn

# Module-level logger

Minor consideration: _resolve_mlflow_resource_arn may raise exceptions (e.g., if the SageMaker API call fails). In the validator, failures are silently swallowed because the validator just returns None. Here, an exception would propagate up and fail evaluate(). Is that the desired behavior? If not, consider wrapping this in a try/except with a warning log, consistent with how the validator handles failures:

if not self.mlflow_resource_arn and self.sagemaker_session:
    try:
        self.mlflow_resource_arn = _resolve_mlflow_resource_arn(self.sagemaker_session)
    except Exception:
        _logger.warning("Failed to resolve MLflow resource ARN during deferred resolution.")

This line is well over 100 characters (the project's line length limit). Please wrap it:

    @patch(
        "sagemaker.train.evaluate.base_evaluator._resolve_mlflow_resource_arn"
    )
    def test_get_base_template_context_deferred_mlflow_resolution(
        self, mock_resolve_mlflow, mock_resolve, mock_session, mock_model_info
    ):

Consider using mock_resolve_mlflow.assert_called_once_with(mock_session) instead of assert_called_with to verify it was called exactly once during _get_base_template_context. Since the mock is also invoked by the validator during construction, you may want to reset the mock after construction (mock_resolve_mlflow.reset_mock()) before calling _get_base_template_context, then assert assert_called_once_with. This makes the test more precise about which call path triggered the resolution.

Setting evaluator.mlflow_resource_arn = None directly after construction to simulate the deferred case works, but it's a bit fragile — it relies on Pydantic allowing direct attribute mutation. A comment explaining why this is necessary (or configuring the mock to return None only during construction) would improve clarity. Actually, you already set mock_resolve_mlflow.return_value = None before construction, so the validator should already leave it as None. Is this explicit assignment redundant? If so, removing it would simplify the test.