Implementation:SeldonIO Seldon core Seldon Model CRD Explainer

Field	Value
Type	Pattern Doc
Overview	Concrete pattern for declaring model explainability configurations in Seldon Core 2 Model CRDs.
Source	samples/models/hf-sentiment-explainer.yaml:L1-9
Domains	MLOps, Explainability
Implements Principle	SeldonIO_Seldon_core_Model_Explainability_Integration
External Dependencies	Kubernetes API, MLServer Alibi-Explain runtime, alibi
Knowledge Sources	Repo (https://github.com/SeldonIO/seldon-core), Doc (https://docs.seldon.io/projects/alibi), Paper (https://arxiv.org/abs/1804.06506)
Last Updated	2026-02-13 00:00 GMT

Code Reference

Sentiment Explainer Model

apiVersion: mlops.seldon.io/v1alpha1
kind: Model
metadata:
  name: sentiment-explainer
spec:
  storageUri: "gs://seldon-models/scv2/samples/mlserver_1.6.0/huggingface/sentiment-explainer"
  explainer:
    type: anchor_text
    pipelineRef: sentiment-explain

Key Parameters

Parameter	Example	Description
metadata.name	sentiment-explainer	Name of the explainer model resource
spec.storageUri	gs://seldon-models/scv2/samples/mlserver_1.6.0/huggingface/sentiment-explainer	Remote URI pointing to the serialized Alibi Explain artifact
spec.explainer.type	anchor_text	Explanation algorithm; anchor_text for NLP models, anchor_tabular for tabular models
spec.explainer.modelRef	(model name, optional)	Name of the base model to explain; used when explaining a single model directly
spec.explainer.pipelineRef	sentiment-explain	Name of the pipeline to use as the black box for generating explanations; used when the base model is accessed through a pipeline

Explainer Types

Type	Use Case	Description
anchor_text	NLP models (sentiment, classification)	Identifies minimal sets of words that guarantee the prediction; uses word replacement perturbations
anchor_tabular	Tabular data models	Identifies feature value ranges that are sufficient for the prediction; uses feature value perturbations

I/O Contract

Inputs

Input	Format	Description
Explainer model artifact	Remote URI	Serialized Alibi Explain instance with pre-configured explanation algorithm parameters
Deployed base model or pipeline	Kubernetes resource	The model or pipeline referenced by modelRef or pipelineRef must be deployed and available

Outputs

Output	Format	Description
Explainer Model resource	Kubernetes Model CRD	Explainer model linked to a base model or pipeline; inference returns anchor explanations
Explanation response	V2 JSON	Anchor text explanation identifying which input features (words) drive the prediction

An example explanation response for a sentiment model:

{
  "outputs": [
    {
      "name": "explanation",
      "datatype": "BYTES",
      "data": ["{\"anchor\": [\"love\"], \"precision\": 0.97, \"coverage\": 0.52}"]
    }
  ]
}

This indicates that the word "love" alone is sufficient (with 97% precision) to produce the POSITIVE classification, and this anchor applies to 52% of similar inputs.

Usage Examples

Deploying the explainer model

seldon model load -f samples/models/hf-sentiment-explainer.yaml
seldon model status sentiment-explainer -w ModelAvailable

Requesting an explanation

seldon model infer sentiment-explainer \
  '{"inputs": [{"name": "args", "shape": [1], "datatype": "BYTES", "data": ["I love this product"]}]}'

Using with a pipeline (speech-to-sentiment)

The explainer is included as a step in the speech-to-sentiment pipeline:

spec:
  steps:
    - name: whisper
    - name: sentiment
      inputs:
      - whisper
      tensorMap:
        whisper.outputs.output: args
    - name: sentiment-input-transform
      inputs:
      - whisper
    - name: sentiment-explainer
      inputs:
      - sentiment-input-transform

In this pipeline, sentiment-explainer receives preprocessed input from sentiment-input-transform and uses the sentiment-explain pipeline (referenced via pipelineRef) as its black box for generating anchor text explanations.

Defining an explainer with modelRef

For simpler cases where the explainer directly references a model (without a pipeline intermediary):

apiVersion: mlops.seldon.io/v1alpha1
kind: Model
metadata:
  name: iris-explainer
spec:
  storageUri: "gs://seldon-models/scv2/samples/mlserver_1.6.0/iris-explainer"
  explainer:
    type: anchor_tabular
    modelRef: iris

Related Pages

• SeldonIO_Seldon_core_Model_Explainability_Integration -- principle that this implementation realizes
• SeldonIO_Seldon_core_Seldon_Pipeline_CRD_Multi_Modal -- integrates with multi-modal pipelines that include explainer steps
• SeldonIO_Seldon_core_Seldon_Model_CRD_HuggingFace -- extends the standard HuggingFace Model CRD with explainer-specific configuration
• SeldonIO_Seldon_core_Seldon_Model_CRD -- specializes the general Model CRD pattern with the spec.explainer section