transformer-lens-interpretability
Find what drives a transformer model's answers.
What it does
Guides investigations that identify which transformer activations and attention heads influence model predictions using TransformerLens.
When to use it
Use when reverse-engineering learned algorithms, studying attention and information flow, or conducting activation patching and circuit analysis.
How to use it
It changes how the agent implements transformer interpretability experiments by defining comparison inputs, caching activations, systematically patching them, and identifying causal hotspots.
What you provide
- An existing project
Uses
What you need · 5
The interpretability examples run as Python code.
Declares transformer-lens version 2.0.0 or newer as a dependency.
Access · 2
This skill
gpt2-small
Loads pretrained GPT-2 weights
transformer-lens
Installs TransformerLens