Seldon Alibi

Seldon Alibi (formally known as Alibi) is an open-source Python library for machine learning model inspection and interpretation, developed and maintained by Seldon Technologies. Part of the broader Seldon ecosystem for ML deployment and serving, Alibi provides a comprehensive collection of algorithms for explaining individual predictions, detecting outliers, and identifying concept drift, making it a versatile toolkit for understanding and monitoring ML model behavior in production. Alibi's explanation capabilities span multiple algorithmic families. For black-box models, the library implements Anchors (rule-based explanations that identify sufficient conditions for a prediction), Contrastive Explanations (CEM), and integrated gradients. For white-box models with gradient access, it offers additional techniques including saliency maps and layer-wise relevance propagation. The library also provides counterfactual explanations, which answer the question 'what would need to change about this input for the model to produce a different prediction?' These counterfactual explanations are increasingly important for regulatory compliance, as they provide actionable recourse information to individuals affected by algorithmic decisions. Beyond local explanations, Alibi includes global explanation methods such as ALE (Accumulated Local Effects) plots and tree-based SHAP implementations. The library also offers outlier detection algorithms including Variational Auto-Encoder based detectors, isolation forests, and Mahalanobis distance detectors, which can identify inputs that fall outside the model's training distribution and may produce unreliable predictions. A key advantage of Alibi is its integration with the Seldon ecosystem. Organizations using Seldon Core or Seldon Deploy for ML model serving can seamlessly add Alibi explainers to their deployment pipelines, enabling real-time explanations for every prediction served. This production-oriented design distinguishes Alibi from research-focused libraries that are primarily used during model development. Alibi supports multiple data types including tabular, text, and image data, and works with models from major frameworks including TensorFlow, PyTorch, and scikit-learn. The library is freely available under the Apache 2.0 license and is actively maintained, though the most advanced deployment and monitoring features require the commercial Seldon Deploy platform.