~/projects/ai-interpretability-journey/

AI Interpretability Journey

Jul 2, 2024 #interpretability#llm#transformerlens

My digital notebook for learning AI interpretability: notes, experiments and insights on making models more transparent.

Purpose

This repository is my research notebook for learning AI interpretability. Experiments look beyond a model’s final output to study how internal Transformer representations and circuits relate to behavior. It brings reading notes together with hands-on notebooks.

What is inside

Notebooks based on ARENA and Transformer Circuits explore tasks such as Find-Min and Find-Median, activations, and component analysis. With TransformerLens I inspect intermediate representations and intervene on activations; visualisations help show which components contribute to an answer.

Scope

This is an evolving collection of learning notes and experiments, not one finished research result. Findings should be read in the context of each notebook’s task and assumptions rather than generalized to all models.

← cd ..