GitHub - TransluceAI/observatory: A toolkit for describing model features and intervening on those features to steer behavior. (original) (raw)

Transluce

Transluce is an independent research lab building open, scalable technology for understanding AI systems and steering them in the public interest.

This repository hosts code for the following projects:

• User Modeling, which is about creating datasets to study language model's latent user representations;
• LatentQA, which trains decoders to read and steer user representations in language models;
• Neuron Descriptions, which automatically generates high-quality descriptions of language model neurons;
• The Monitor interface, which helps humans observe, understand, and steer the internal computations of language models.

Table of Contents

• Installation
• Getting Started
  • User Modeling
  • LatentQA
  • Neuron Descriptions
  • Monitor
• Using luce
  • Package and environment management
  • Using Jupyter notebooks
• Citation

Installation

First clone this repo:

git clone https://github.com/TransluceAI/observatory.git

Installing luce

Next, we'll install luce, a command-line tool that manages project environments and dependencies. It will significantly simplify setup for downstream projects.

To install luce, add the following to your shell profile (e.g., .bashrc, .zshrc):

... existing shell config

export TRANSLUCE_HOME= source $TRANSLUCE_HOME/lib/lucepkg/scripts/shellenv.sh

Make sure to source your shell profile:

source ~/.bashrc # for bash users

OR

source ~/.zshrc # for zsh users

Then run:

luce uv install # install uv package manager luce install # install base environment

to install the uv package manager and base virtual environment, respectively. The base venv includes basic packages like the Jupyter kernel, pre-commit, etc.

Setting up environment variables

Finally, clone the .env.template file to .env and fill in the missing values.

These variables are always required:

• OPENAI_API_KEY / OPENAI_API_ORG: OpenAI API key and organization ID.
• ANTHROPIC_API_KEY: Anthropic API key.
• HF_TOKEN: Required for accessing gated models (e.g., Llama-3.1) on HuggingFace.

The rest of the variables are only required for running the NeuronDB or Monitor; you can safely ignore them for now.

Getting Started

User Modeling

See the user modeling README for generating data to study how language models represent users.

LatentQA

See the latentqa README for training LatentQA decoders to read and steer user representations in language models.

Neuron Descriptions

See the description generation README for generating neuron descriptions automatically.

Monitor

See the Monitor README for instructions on how to set up a local development environment.

Using luce

Package and environment management

Each folder under lib/ and project/ has its own venv. Use luce to:

Install all dependencies

luce install --all

Install and activate a specific package

luce install

Activate a venv and cd to its directory

luce activate

Deactivate the current package

deactivate

You may need to use --force to reinstall a package that already exists; this removes poetry.lock:

luce install --force

Using Jupyter notebooks

We include utilities to register Jupyter kernels into the top-level environment. To register a kernel for a package, run:

luce nb register

To start a notebook server that can call any of the registered kernels, run:

luce nb start --port

You'll get a readout of the notebook server URL, which you can use to connect to the notebook server via the web or an IDE.

Support

If you run into any issues, please file an issue or reach out to us at info@transluce.org. We're happy to help!

Citation

If you use this code for your research, please cite our work:

@misc{choi2025scalably, author = {Choi, Dami and Huang, Vincent and Schwettmann, Sarah and Steinhardt, Jacob}, title = {Scalably Extracting Latent Representations of Users}, year = {2025}, month = {November}, day = {25}, howpublished = {\url{https://transluce.org/user-modeling}} }

@misc{choi2024automatic, author = {Choi, Dami and Huang, Vincent and Meng, Kevin and Johnson, Daniel D and Steinhardt, Jacob and Schwettmann, Sarah}, title = {Scaling Automatic Neuron Description}, year = {2024}, month = {October}, day = {23}, howpublished = {\url{https://transluce.org/neuron-descriptions}} }