transformer-lens-interpretability

What problem does it solve?

This Skill provides tools and guidance for understanding the internal workings of transformer models, enabling researchers to reverse-engineer learned algorithms and analyze model behavior.

Core Features & Use Cases

• Mechanistic Interpretability: Deep dive into transformer circuits, attention patterns, and activation flows.
• Activation Patching: Perform causal interventions to understand the impact of specific activations on model output.
• Use Case: A researcher wants to understand how a language model identifies and resolves pronoun coreferences. They can use this Skill to isolate the specific attention heads and layers responsible for this task by patching activations between different input prompts.

Quick Start

Use the transformer-lens-interpretability skill to analyze attention patterns in layer 3 of a GPT-2 small model.