Exploratory Analysis Demo - TransformerLens Documentation

This notebook demonstrates how to use the TransformerLens library to perform exploratory analysis. The notebook tries to replicate the analysis of the Indirect Object Identification circuit in the Interpretability in the Wild paper. If running in Google Colab, go to Runtime > Change Runtime Type and select GPU as the hardware accelerator. Look up unfamiliar terms in the mech interp explainer You can run all this code for yourself The graphs are interactive Use the table of contents pane in the sidebar to navigate (in Colab) or VSCode’s “Outline” in the explorer tab. Collapse irrelevant sections with the dropdown arrows Search the page using the search in the sidebar (with Colab) not CTRL+F You can ignore this part: It’s just for use internally to setup the tutorial in different environments. You can delete this section if using in your own repo. We turn automatic differentiation off, to save GPU memory, as this notebook focuses on model inference not model training. Some plotting helpe

Exploratory Analysis Demo ¶ This notebook demonstrates how to use the TransformerLens library to perform exploratory analysis. The notebook tries to replicate the analysis of the Indirect Object Identification circuit in the Interpretability in the Wild paper. Tips for Reading This ¶ If running in Google Colab, go to Runtime > Change Runtime Type and select GPU as the hardware accelerator. Look up unfamiliar terms in the mech interp explainer You can run all this code for yourself The graphs are interactive Use the table of contents pane in the sidebar to navigate (in Colab) or VSCode’s “Outli

related reading

• A Mathematical Framework for Transformer Circuitstransformer-circuits.pub
• Chapter 1: Transformer Interpretability - ARENAlearn.arena.education
• Transformer Explainer: LLM Transformer Model Visually Explainedpoloclub.github.io
• Transformer Circuits Threadtransformer-circuits.pub
• Circuit Tracing: Revealing Computational Graphs in Language Modelstransformer-circuits.pub
• Chapter 1: Transformer Interpretability - ARENAlearn.arena.education
• Attribution Patching: Activation Patching At Industrial Scale - Neel Nandaneelnanda.io
• Chapter 1: Transformer Interpretability - ARENAlearn.arena.education
• Chapter 1: Transformer Interpretability - ARENAlearn.arena.education
• An Extremely Opinionated Annotated List of My Favourite Mechanistic Interpretability Papers v2 — AI Alignment Forumalignmentforum.org
• The Annotated Transformernlp.seas.harvard.edu
• Gears-Level Mental Models of Transformer Interpretability — LessWronglesswrong.com