cdpierse/transformers-interpret

Supports attribution-based explainability for both text and vision transformers through gradient-based methods (Integrated Gradients), generating token-level importance scores. Provides multiple explainer classes tailored to different task types—sequence classification, pairwise classification, question answering, and image classification—with built-in visualization as interactive HTML or static PNG outputs. Integrates directly with Hugging Face transformers' model and tokenizer APIs, supporting any pretrained or fine-tuned model from the ecosystem.

1,413 stars. No commits in the last 6 months.