mechlens

Personal mechanistic-interpretability toolkit.

A HookedTransformer wrapper for GPT-style language models that exposes every internal activation (residual stream, attention heads, MLPs) via named hooks. Loads 9,000+ open-source models across 50+ architecture families and lets you cache, edit, remove, or replace activations as the model runs.

Install

pip install -e .

Quickstart

from transformer_lens import HookedTransformer

model = HookedTransformer.from_pretrained("gpt2")
logits, cache = model.run_with_cache("The capital of France is")
print(cache["blocks.0.attn.hook_pattern"].shape)

See demos/Main_Demo.ipynb for the full walkthrough.

Docs

Docs source lives in docs/. Build locally with:

cd docs && make html

License

MIT. See LICENSE.