Inseq is a Python toolkit for post-hoc interpretability analysis of sequence generation models, targeting researchers and practitioners in NLP. It democratizes access to various attribution methods, enabling deeper understanding of model behavior and facilitating reproducible research.

How It Works

Inseq integrates with Hugging Face Transformers, supporting both encoder-decoder and decoder-only architectures. It implements a wide range of attribution methods, including gradient-based (e.g., Integrated Gradients, DeepLIFT), attention-based, and perturbation-based techniques. The library allows for flexible post-processing of attribution maps via Aggregator classes and supports custom attribution targets using "step functions" to extract scores like logits, probabilities, or entropy at each generation step.

Quick Start & Requirements

• Install latest stable version: pip install inseq
• Install extras for visualization: pip install inseq[notebook,datasets]
• Requires Python >= 3.10, <= 3.12.
• tokenizers installation requires a Rust compiler.
• sentencepiece installation may require cmake, build-essential, pkg-config.
• Official Docs: https://inseq.org
• Tutorial: https://github.com/inseq-team/inseq/blob/main/examples/inseq_tutorial.ipynb

Highlighted Details

• Supports a broad spectrum of attribution methods, extending Captum's capabilities.
• Offers visualization in notebooks, browsers, and the command line.
• Includes a CLI for batch attribution on datasets and context dependence analysis.
• Enables custom attribution targets and extraction of intermediate generation scores.

Maintenance & Community

• Active development with a clear roadmap.
• Community support via Discord: https://discord.gg/V5VgwwFPbu
• Active presence on Twitter: https://twitter.com/InseqLib

Licensing & Compatibility

• MIT License. Permissive for commercial use and integration with closed-source projects.

Limitations & Caveats

• Python version compatibility is restricted to 3.10-3.12.
• Installation of certain dependencies (tokenizers, sentencepiece) may require additional system-level setup.