AI Interpretability Research Platform

Use Workbench now!

GitHub | Discord | Forum | Twitter

---

About

Workbench is a UI for doing exploratory analysis on open source AI models by applying interpretability techniques. It leverages both NNsight and NDIF to provide an interactive environment for exploring LLM internals and building experiments.

---

Setup

Requirements

1. Node.js >= 18.18.0
2. Install uv
3. Install bun

Steps

1. Create a venv using uv venv
2. Activate it afterwards using the printed command
3. Run uv sync --extra dev
4. Setup .env and workbench/_api/.env following their respective .env.template files.

Now, run the frontend and backend together, with:

1. bash ./scripts/web.sh
2. bash ./scripts/api.sh

---

Local Database

To use a local SQLite database instead of Supabase, add these to your root .env:

NEXT_PUBLIC_LOCAL_DB=true
LOCAL_SQLITE_URL=./local.db

Then create the database tables:

cd workbench/_web
bunx drizzle-kit generate
bunx drizzle-kit push