Interpreto
Maintainer:IRT-St-Exupery
Description
Interpreto is a modular Python interpretability toolbox for explaining Large Language Models (LLMs). It provides attribution methods (KernelShap, LIME, Integrated Gradients, Occlusion, Sobol and more), concept-based methods (supervised probes, unsupervised dictionary learning such as sparse autoencoders, NMF, ICA, PCA), and evaluation metrics (Insertion/Deletion, ConSim, sparsity), covering both classification and text-generation models.
Owner:IRT-St-Exupery
Keywords:explainabilityinterpretability
Understanding why a Large Language Model produces a given prediction or generation is increasingly required as LLMs are adopted in higher-stakes settings, for both auditability and end-user trust.
Interpreto brings attribution methods (gradient-based and inference-based), concept-based methods (supervised and unsupervised), and their evaluation metrics together in one modular framework built on HuggingFace transformers and PyTorch, covering both classification (`...ForSequenceClassification`) and text-generation (`...ForCausalLM`) models.
Used by ML-Algorithm Engineers who need to explain or audit an LLM's predictions or generations, compare attribution/concept-based methods against each other, or evaluate explanation quality itself (via the Insertion/Deletion, ConSim and sparsity metrics).
