Implementation:SeldonIO Seldon core Seldon Model Infer Explainer

Field	Value
Type	Wrapper Doc
Overview	Concrete CLI tool for generating explanations by sending inference requests to explainer endpoints in Seldon Core 2.
Domains	Explainability, Inference
Workflow	Model_Explainability
Related Principle	SeldonIO_Seldon_core_Explanation_Generation
Source	samples/explainer-examples.md:L65-146, docs-gb/cli/seldon_model_infer.md:L1-35
Last Updated	2026-02-13 00:00 GMT

Code Reference

CLI Command Syntax

Source: docs-gb/cli/seldon_model_infer.md:L1-35

seldon model infer <model-name> <json-payload> [flags]

Tabular Explanation Request

Source: samples/explainer-examples.md:L65-100

seldon model infer income-explainer \
  '{"inputs": [{"name": "predict", "shape": [1, 12], "datatype": "FP32", "data": [[39,7,1,1,1,1,4,1,2174,0,40,9]]}]}'

Text Explanation Request

Source: samples/explainer-examples.md:L105-146

seldon model infer sentiment-explainer \
  '{"inputs": [{"name": "predict", "shape": [1], "datatype": "BYTES", "data": ["a]"}]}'

Key Parameters

Parameter	Description	Example
explainerName	Name of the deployed explainer model	income-explainer, sentiment-explainer
inputs[].name	Input tensor name	predict
inputs[].shape	Tensor shape	[1, 12] for tabular, [1] for text
inputs[].datatype	Data type	FP32 for numeric features, BYTES for text
inputs[].data	Input feature values	39,7,1,1,1,1,4,1,2174,0,40,9 or ["a]"

I/O Contract

Inputs

• V2 JSON payload: A JSON object conforming to the V2 inference protocol, with input features matching the base model's expected format:
  • Tabular (FP32): Numeric array with shape matching the model's feature count
  • Text (BYTES): String array with the text to explain

Outputs

• V2 response with explanation metadata:
  • anchor: List of anchor features/conditions that guarantee the prediction (e.g., ["Marital Status = Separated", "Age > 37"])
  • precision: Precision score of the anchor (fraction of perturbations where prediction holds when anchor conditions are met)
  • coverage: Coverage metric (fraction of the dataset where the anchor applies)
  • raw.feature: Arrays with covered_true and covered_false examples showing perturbation results
  • meta.params: Explainer configuration parameters used during explanation generation

External Dependencies

• seldon CLI
• curl (alternative to CLI)
• V2 inference protocol
• Alibi-Explain runtime on MLServer

Usage Examples

Generating a Tabular Explanation

# Request an explanation for an income prediction
seldon model infer income-explainer \
  '{"inputs": [{"name": "predict", "shape": [1, 12], "datatype": "FP32", "data": [[39,7,1,1,1,1,4,1,2174,0,40,9]]}]}'

# Example response (truncated) includes:
# - anchor: ["Marital Status = Separated", "Capital Gain <= 0"]
# - precision: 0.97
# - coverage: 0.42

Generating a Text Explanation

# Request an explanation for a sentiment prediction
seldon model infer sentiment-explainer \
  '{"inputs": [{"name": "predict", "shape": [1], "datatype": "BYTES", "data": ["a]"}]}'

# Example response includes:
# - anchor: words whose presence anchors the prediction
# - precision: probability the prediction holds with anchor words
# - coverage: fraction of texts where the anchor applies

Using curl as an Alternative

# Direct HTTP request to the explainer endpoint
curl -X POST http://<inference-gateway>/v2/models/income-explainer/infer \
  -H "Content-Type: application/json" \
  -d '{"inputs": [{"name": "predict", "shape": [1, 12], "datatype": "FP32", "data": [[39,7,1,1,1,1,4,1,2174,0,40,9]]}]}'

Knowledge Sources

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

Related Pages

• SeldonIO_Seldon_core_Explanation_Generation - parent principle - Generating human-interpretable explanations for individual predictions.
• SeldonIO_Seldon_core_Seldon_Model_Status_Explainer - prerequisite - The explainer must be confirmed ready before sending requests.
• SeldonIO_Seldon_core_Alibi_Explainer_Training - related - How the explainer artifacts were trained.
• SeldonIO_Seldon_core_Seldon_Pipeline_CRD_Explainer - alternative - Explanation via pipeline integration.