crosscoders

A Python package to run crosscoders.

Table of Contents

1. Installation
2. Configuration
   • Environment Variables
   • Run Configs
3. Usage

Installation

1. Install conda

2. conda env create --file=env.yml

Configuration

Environment Variables

Create a .env file in the repo root directory and add the following lines, modifying PROJECT_ROOT_DIR and CONFIG_FILEPATH to point to the absolute path of the repo root directory and the run config file you want to use.

export PROJECT_ROOT_DIR="."
export CONFIG_FILEPATH="${PROJECT_ROOT_DIR}/src/scripts/configs/cfg.yml"   # eg

See src/crosscoders/constants.py for a list of REQUIRED_ENV_VARS.

See src/scripts/configs/ for preset run configurations.

Run Configs

The run config files are the primary place to tweak anything in the code. The syntax is defined by the config dataclasses at src/crosscoders/dataclasses/configs/.

Currently, there are two base config objects that contain all others: GlobalsConfig and RunnerConfig. Your cfg.yml should have only two root dictionaries: GLOBALS and RUNNER.

For the most part, the field names in the cfg.yml are uppercase versions of the config dataclass names minus Config.

See src/crosscoders/dataclasses/configs/.

Usage

This package is intended to be used as a CLI called xc, although it can be used programmatically as well.

There is only one required argument, mode, which specifies which sub-script to run.

See src/scripts/main.py.

# export CONFIG_FILEPATH="${PROJECT_ROOT_DIR}/src/scripts/configs/data.yml
xc data

# export CONFIG_FILEPATH="${PROJECT_ROOT_DIR}/src/scripts/configs/train.yml
xc train