Principle:SeldonIO Seldon core Explainer Model Deployment
| Field | Value |
|---|---|
| Overview | Declaring and deploying explainer models with explicit links to their base models or pipelines. |
| Domains | MLOps, Explainability, Kubernetes |
| Workflow | Model_Explainability |
| Related Implementation | SeldonIO_Seldon_core_Seldon_Model_CRD_Explainer_Deploy |
| Last Updated | 2026-02-13 00:00 GMT |
Description
Seldon Core 2 explainer models use the Model CRD with an additional spec.explainer section. This section specifies the explainer algorithm type (anchor_tabular, anchor_text, kernel_shap) and links to the base model via modelRef (for direct model explanation) or pipelineRef (for pipeline-level explanation). The explainer artifact is served by MLServer's Alibi-Explain runtime.
The key fields in the spec.explainer section are:
- type: The Alibi explanation algorithm to use. Determines how perturbations are generated and how explanations are constructed.
- modelRef: References a deployed Model by name. The explainer queries this model's inference endpoint for black-box predictions during explanation generation.
- pipelineRef: References a deployed Pipeline by name. Used when the explainer should explain the output of a multi-step inference pipeline rather than a single model.
The storageUri field points to the serialized explainer artifact (trained via the alibi library), which MLServer loads using the Alibi-Explain runtime.
Theoretical Basis
The modelRef/pipelineRef linkage creates a dependency graph where the explainer knows which model to query for black-box predictions. The type field determines which Alibi algorithm the runtime will use to generate explanations. This decoupling allows different explainer types to be swapped without changing the base model.
The separation of the explainer from the base model follows the model-agnostic explanation paradigm: the explainer only needs access to the model's prediction function, not its internal structure. This means:
- Any classifier conforming to the V2 inference API can be explained.
- Explainers can be updated or replaced independently of the base model.
- Multiple explainer types can be deployed for the same base model simultaneously.
Usage
When deploying an explainer model to provide interpretable explanations for an existing classifier. The base model (referenced by modelRef) or pipeline (referenced by pipelineRef) must already be deployed and available.
Knowledge Sources
- Repo: SeldonIO/seldon-core
- Doc: Seldon Core 2 Documentation
Related Pages
- SeldonIO_Seldon_core_Seldon_Model_CRD_Explainer_Deploy - implements this principle - Concrete pattern for declaring and deploying explainer models.
- SeldonIO_Seldon_core_Explainer_Base_Model_Deployment - prerequisite - The base model must be deployed first.
- SeldonIO_Seldon_core_Explainer_Readiness_Verification - next step - Verifying the explainer is ready after deployment.
- SeldonIO_Seldon_core_Explainer_Pipeline_Integration - related principle - Using pipelineRef for pipeline-level explanation.
Implementation:SeldonIO_Seldon_core_Seldon_Model_CRD_Explainer_Deploy