Principle:SeldonIO Seldon core Explainer Base Model Deployment

Field	Value
Overview	Deploying the base classifier model that the explainer will generate explanations for.
Domains	MLOps, Explainability
Workflow	Model_Explainability
Related Implementation	SeldonIO_Seldon_core_Seldon_Model_Load_For_Explainer
Last Updated	2026-02-13 00:00 GMT

Description

An explainer requires a deployed base model to function as its black-box predictor. The base model must be loaded and available before the explainer model is deployed, since the explainer references the base model via modelRef and uses it for perturbation-based explanation generation.

In Seldon Core 2, the base model is deployed using the standard Model CRD and loaded via the seldon model load CLI command. The model must reach the ModelAvailable condition before the explainer can be deployed. This ordering is critical because:

• The explainer's modelRef field creates a runtime dependency on the base model.
• During explanation generation, the explainer sends hundreds or thousands of perturbed inputs to the base model for prediction.
• If the base model is unavailable, explanation requests will fail.

Theoretical Basis

Black-box explanations require access to the model's prediction function. In Seldon Core 2, the explainer queries the base model's inference endpoint to evaluate perturbed inputs. This separation allows the same explainer to work with any model that conforms to the V2 inference API.

The perturbation-based approach means the explainer treats the base model as an oracle:

• AnchorTabular: Perturbs feature values according to the learned discretization and queries the base model for each perturbation.
• AnchorText: Replaces words with UNK tokens or similar words and queries the base model.
• KernelShap: Masks subsets of features and queries the base model to compute marginal contributions.

This design decouples the explainer from the model implementation, allowing any classifier conforming to the V2 inference protocol to be explained.

Usage

Before deploying an explainer model in Seldon Core 2. The base model must be loaded and confirmed as ModelAvailable before the explainer model is loaded.

Knowledge Sources

• Repo: SeldonIO/seldon-core
• Doc: Seldon Core 2 Documentation

Related Pages

• SeldonIO_Seldon_core_Seldon_Model_Load_For_Explainer - implements this principle - Concrete CLI tool for deploying base classifier models as prerequisites for explainers.
• SeldonIO_Seldon_core_Explainer_Model_Deployment - next step - Deploying the explainer model once the base model is ready.
• SeldonIO_Seldon_core_Explainer_Model_Training - prerequisite - Training the explainer model artifacts before deployment.

Implementation:SeldonIO_Seldon_core_Seldon_Model_Load_For_Explainer