repana 0.0.3

A simple framework for analyzing model representations

Table of Contents

• Overview
• Features
• Installation
• Quick Start
• Acknowledgements

Overview

Repana was developed based on the repeng library for research purposes. It is intended for doing representation analysis of large language models. Repana allows you to insert a (or more) ControlBlock into a standard torch transformer model creating a ControlModel instance. You can then "read" model representations, use these to derive a ControlVector, which can be applied to the ControlModel during inference.

Features

• ControlModel
• ControlVector
  • Reading vector
  • Contrast vector
  • PCA-based contrast vector

Installation

pip install repana

Acknowledgements

This library is based off of the repeng library (MIT License) developed by Theia Vogel. If you want to merely train control vectors fast, we suggest taking a look at that great library!