TransformerLens in AI answers

TransformerLens is a Python library designed for mechanistic interpretability of generative language models, enabling users to reverse-engineer the algorithms learned by models like GPT-2 from their weights. It can load 15,000+ open-source language models across 140+ architecture families, exposes and caches internal activations, and provides capabilities to edit, remove, or replace activations during inference. It provides a TransformerBridge interface to run models and obtain logits and activations, supports HuggingFace weights with compatibility options, and is maintained by researchers Bryce Meyer, Jonah Larson, and Neel Nanda.