MaxBosch260 / IEinAI- Public

forked from UFO-101/auto-circuit

Notifications You must be signed in to change notification settings
Fork 0
Star 0

Repository for investigating circuits in multilingual language models for the subject-verb agreement task.

ufo-101.github.io/auto-circuit

0 stars 13 forks Branches Tags Activity

Notifications

Name		Name	Last commit message	Last commit date
Latest commit History 68 Commits
.github		.github
.measurement_cache		.measurement_cache
.prune_scores_cache		.prune_scores_cache
.vscode		.vscode
auto_circuit		auto_circuit
data		data
datasets		datasets
docs		docs
experiments		experiments
tests		tests
.dockerignore		.dockerignore
.gitignore		.gitignore
.pre-commit-config.yaml		.pre-commit-config.yaml
Dockerfile		Dockerfile
README.md		README.md
Transformer Circuit Metrics are not Robust.pdf		Transformer Circuit Metrics are not Robust.pdf
mkdocs.yml		mkdocs.yml
poetry.lock		poetry.lock
poetry.toml		poetry.toml
pyproject.toml		pyproject.toml
test.job		test.job

Repository files navigation

AutoCircuit

A library for efficient patching and automatic circuit discovery.

Read the draft paper

Transformer Circuit Metrics are not Robust

Getting Started

pip install auto-circuit

Easy and Efficient Edge Patching

https://github.com/UFO-101/auto-circuit/blob/03ce552ccdea98fc6fed3b05cfb2df9b0eb4e323/experiments/demos/zero_ablate_an_edge.py#L20-L26

Different Ablation Methods

https://github.com/UFO-101/auto-circuit/blob/03ce552ccdea98fc6fed3b05cfb2df9b0eb4e323/experiments/demos/patch_an_edge.py#L37

Automatic Circuit Discovery

https://github.com/UFO-101/auto-circuit/blob/03ce552ccdea98fc6fed3b05cfb2df9b0eb4e323/experiments/demos/patch_an_edge.py#L49-L55

Visualization

https://github.com/UFO-101/auto-circuit/blob/5777057b445d55b7d4695eed8aface9d6bcefbe3/experiments/demos/patch_some_edges.py#L49

About

Repository for investigating circuits in multilingual language models for the subject-verb agreement task.

UFO-101.github.io/auto-circuit

Report repository

Releases

No releases published

Packages

No packages published

Languages

Python 99.6%
Other 0.4%