Implementation:SeldonIO Seldon core Seldon Pipeline CRD Explainer

Field	Value
Type	Pattern Doc
Overview	Concrete pattern for integrating explainers into Seldon Core 2 pipelines with transform steps.
Domains	Explainability, Data_Flow
Workflow	Model_Explainability
Related Principle	SeldonIO_Seldon_core_Explainer_Pipeline_Integration
Source	samples/pipelines/sentiment-explain.yaml:L1-16, samples/models/hf-sentiment-explainer.yaml:L1-9
Last Updated	2026-02-13 00:00 GMT

Code Reference

Sentiment Explain Pipeline

Source: samples/pipelines/sentiment-explain.yaml:L1-16

apiVersion: mlops.seldon.io/v1alpha1
kind: Pipeline
metadata:
  name: sentiment-explain
spec:
  steps:
    - name: sentiment
    - name: sentiment-output-transform
      inputs:
      - sentiment
      tensorMap:
        sentiment-explain.inputs.predict: array_inputs
  output:
    steps:
    - sentiment-output-transform

Explainer with pipelineRef

Source: samples/models/hf-sentiment-explainer.yaml:L1-9

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	Description	Example
spec.steps[].name	Pipeline step name (references a deployed model)	sentiment, sentiment-output-transform
spec.steps[].inputs	Input dependencies from other steps	- sentiment
spec.steps[].tensorMap	Maps input tensor names to the step's expected names	sentiment-explain.inputs.predict: array_inputs
spec.output.steps	Final output step(s) of the pipeline	- sentiment-output-transform
spec.explainer.pipelineRef	References the explanation pipeline for black-box inference	sentiment-explain
Custom transform model	MLServer model extending MLModel to reshape tensors	SentimentOutputTransformRuntime

I/O Contract

Inputs

• Deployed sentiment-explain pipeline: The Pipeline CRD must be loaded with all steps (base model and transform models) available
• Custom output transform model: A deployed model that reshapes the pipeline's output to match the explainer's expected format
• Explainer with pipelineRef: The explainer CRD references the pipeline by name

Outputs

• Pipeline CRD: sentiment-explain: A deployed pipeline that routes inference through the base model and output transform
• Explainer references pipeline via pipelineRef: The explainer sends its black-box queries through the full pipeline for explanation generation

External Dependencies

• Kubernetes API
• Kafka (for pipeline step communication)
• Custom MLServer transform models (extending MLModel)
• Seldon scheduler

Usage Examples

Full Pipeline Explainer Deployment

# Step 1: Load the base sentiment model
seldon model load -f samples/models/huggingface-sentiment.yaml
seldon model status sentiment -w ModelAvailable

# Step 2: Load the output transform model
seldon model load -f samples/models/sentiment-output-transform.yaml
seldon model status sentiment-output-transform -w ModelAvailable

# Step 3: Create the explanation pipeline
seldon pipeline load -f samples/pipelines/sentiment-explain.yaml
seldon pipeline status sentiment-explain -w PipelineReady

# Step 4: Load the explainer with pipelineRef
seldon model load -f samples/models/hf-sentiment-explainer.yaml
seldon model status sentiment-explainer -w ModelAvailable

# Step 5: Request an explanation
seldon model infer sentiment-explainer \
  '{"inputs": [{"name": "predict", "shape": [1], "datatype": "BYTES", "data": ["This movie was great"]}]}'

Transform Model Pattern

Custom transform models extend MLServer's MLModel class to reshape tensors between the pipeline's format and the explainer's expected format:

from mlserver import MLModel
from mlserver.types import InferenceRequest, InferenceResponse

class SentimentOutputTransformRuntime(MLModel):
    async def load(self) -> bool:
        # Initialize any required state
        return True

    async def predict(self, payload: InferenceRequest) -> InferenceResponse:
        # Reshape the sentiment model's output tensor
        # to match the explainer's expected input format
        ...

Pipeline tensorMap Explanation

The tensorMap field remaps tensor names between pipeline steps:

tensorMap:
  sentiment-explain.inputs.predict: array_inputs

This maps the pipeline's input tensor sentiment-explain.inputs.predict to the transform model's expected input tensor name array_inputs, enabling seamless data flow without modifying the transform model's code.

Knowledge Sources

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

Related Pages

• SeldonIO_Seldon_core_Explainer_Pipeline_Integration - parent principle - Integrating explainer models into Seldon pipelines.
• SeldonIO_Seldon_core_Seldon_Model_CRD_Explainer_Deploy - related - Direct model-level explainer deployment (using modelRef instead of pipelineRef).
• SeldonIO_Seldon_core_Seldon_Model_Infer_Explainer - next step - Sending explanation requests to the pipeline-integrated explainer.