Principle:SeldonIO Seldon core Model Explainability Integration

Field	Value
Overview	Adding model explainability capabilities to deployed models using Alibi Explain integration.
Domains	MLOps, Explainability
Related Implementation	SeldonIO_Seldon_core_Seldon_Model_CRD_Explainer
Knowledge Sources	Repo (https://github.com/SeldonIO/seldon-core), Doc (https://docs.seldon.io/projects/alibi), Paper (Anchors: https://arxiv.org/abs/1804.06506)
Last Updated	2026-02-13 00:00 GMT

Description

Seldon Core 2 supports explainability through the Model CRD's spec.explainer section. The explainer field links an explanation model to a base model (via modelRef) or pipeline (via pipelineRef). The explainer type specifies the algorithm (e.g., anchor_text, anchor_tabular). The explainer model artifact is served by MLServer's Alibi-Explain runtime.

The explainability integration involves three components:

1. Explainer model artifact -- a serialized Alibi Explain instance stored at a remote URI, loaded by the Alibi-Explain MLServer runtime
2. Base model or pipeline reference -- the modelRef or pipelineRef field that tells the explainer which model to explain (the black box)
3. Explainer type -- the algorithm used for generating explanations (e.g., anchor_text for NLP models, anchor_tabular for tabular data models)

When an inference request is sent to an explainer model, the Alibi-Explain runtime:

1. Receives the input data
2. Calls the base model/pipeline as a black box to get predictions
3. Applies the explanation algorithm (e.g., anchor perturbation) to identify decision-relevant features
4. Returns the explanation alongside or instead of the raw prediction

Theoretical Basis

Model explainability provides human-interpretable explanations for predictions. Anchor explanations identify minimal sufficient conditions ("anchors") that guarantee the same prediction. The explainer uses the base model as a black box, perturbing inputs to find the decision boundary.

The Anchor algorithm (Ribeiro et al., 2018) works as follows:

1. Perturbation -- The algorithm generates perturbed versions of the input by randomly replacing features (words for text, values for tabular data).
2. Precision estimation -- For each candidate anchor (subset of features), the algorithm estimates the probability that fixing those features maintains the original prediction.
3. Beam search -- A beam search algorithm efficiently explores the space of possible anchors, adding features one at a time.
4. Sufficient condition -- The algorithm terminates when it finds a minimal set of features (the anchor) whose precision exceeds a configurable threshold (typically 95%).

For text models (anchor_text), anchors identify the minimal set of words in the input that, when present, guarantee the same classification regardless of the surrounding context.

For tabular models (anchor_tabular), anchors identify feature value ranges that are sufficient for the prediction.

Usage

This principle applies when adding interpretability to deployed models or pipelines in production, including:

• Explaining sentiment analysis predictions by identifying which words drive the classification
• Explaining tabular model predictions by identifying which feature values are decision-relevant
• Integrating explainability into multi-modal pipelines as a branch step
• Meeting regulatory or compliance requirements for model transparency

Related Pages

• SeldonIO_Seldon_core_Seldon_Model_CRD_Explainer -- implements this principle with concrete Model CRD YAML for explainer configuration
• SeldonIO_Seldon_core_Multi_Modal_Pipeline_Composition -- integrates with pipeline composition where explainers are pipeline branches
• SeldonIO_Seldon_core_Seldon_Pipeline_CRD_Multi_Modal -- used by the speech-to-sentiment pipeline which includes an explainer step
• SeldonIO_Seldon_core_HuggingFace_Model_Resource_Definition -- extends the Model CRD pattern with explainer-specific fields

Implementation:SeldonIO_Seldon_core_Seldon_Model_CRD_Explainer