![[2512.09730] Interpreto: An Explainability Library for Transformers](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2512.09730] Interpreto: An Explainability Library for Transformers
Summary
Interpreto is an open-source library designed for interpreting HuggingFace transformers, offering both attribution methods and concept-based explanations for enhanced model explainability.
Why It Matters
As AI models become increasingly complex, understanding their decision-making processes is crucial for trust and accountability. Interpreto provides tools that bridge research and practical applications, making it easier for developers and researchers to interpret transformer models effectively.
Key Takeaways
- Interpreto offers a unified API for interpreting HuggingFace models.
- The library includes both attribution methods and concept-based explanations.
- It features an end-to-end pipeline for concept learning and interpretation.
- Interpreto supports both classification and text generation tasks.
- The library aims to enhance transparency and understanding of AI models.
Computer Science > Computation and Language arXiv:2512.09730 (cs) [Submitted on 10 Dec 2025 (v1), last revised 22 Feb 2026 (this version, v2)] Title:Interpreto: An Explainability Library for Transformers Authors:Antonin Poché, Thomas Mullor, Gabriele Sarti, Frédéric Boisnard, Corentin Friedrich, Charlotte Claye, François Hoofd, Raphael Bernas, Céline Hudelot, Fanny Jourdan View a PDF of the paper titled Interpreto: An Explainability Library for Transformers, by Antonin Poch\'e and 9 other authors View PDF HTML (experimental) Abstract:Interpreto is an open-source Python library for interpreting HuggingFace language models, from early BERT variants to LLMs. It provides two complementary families of methods: attribution methods and concept-based explanations. The library bridges recent research and practical tooling by exposing explanation workflows through a unified API for both classification and text generation. A key differentiator is its end-to-end concept-based pipeline (from activation extraction to concept learning, interpretation, and scoring), which goes beyond feature-level attributions and is uncommon in existing libraries. Comments: Subjects: Computation and Language (cs.CL); Machine Learning (cs.LG) ACM classes: I.2.7 Cite as: arXiv:2512.09730 [cs.CL] (or arXiv:2512.09730v2 [cs.CL] for this version) https://doi.org/10.48550/arXiv.2512.09730 Focus to learn more arXiv-issued DOI via DataCite Submission history From: Antonin Poché [view email] [v1] Wed, 10 Dec ...
