Glossary · Term

TransformerLens

← all terms

Definition

Plain language

A popular Python library that lets researchers peek inside transformer models layer by layer.

As stated in the literature

An open-source library providing hooks and tooling for inspecting and intervening on internal activations of transformer language models.

Why it matters: It lowers the barrier to mechanistic interpretability work, letting more researchers probe and intervene on real model internals.

For example, a researcher can use TransformerLens to hook into a specific attention head and zero out its contribution to see how the model's answer changes.

Heard on the show

“They load it through TransformerLens, a tool that lets you intercept and modify the model's internal state at every layer of the forward pass.”
Episode 018 — Language Models Compute the Rational Move, Then Override It

Mentioned in 1 episode

  1. 018
    Language Models Compute the Rational Move, Then Override It

Related terms