Open Concept Steering

Open Concept Steering is an open-source library for discovering and manipulating interpretable features in large language models using Sparse Autoencoders (SAEs). Inspired by Anthropic's work on Scaling Monosemanticity and Golden Gate Claude, this project aims to make concept steering accessible to the broader research community.

Features

Coming soon!

• Universal Model Support: Train SAEs on any HuggingFace transformer model
• Feature Discovery: Find interpretable features representing specific concepts
• Concept Steering: Amplify or suppress discovered features to influence model behavior
• Interactive Chat: Chat with models while manipulating their internal features

Pre-trained Models

In the spirit of fully open-source models, we have started training SAEs on OLMo 2 7B.

We provide pre-trained SAEs and discovered features for popular models on HuggingFace:

Each model repository will include:

• Trained SAE weights
• Catalog of discovered interpretable features
• Example steering configurations

Quick Start

Examples (In progress)

See the examples/ directory for detailed notebooks demonstrating:

• Training SAEs on different models
• Finding and analyzing features
• Steering model behavior
• Interactive chat sessions

License

This project is licensed under the MIT License.

Citation

If you feel compelled to cite this library in your work, feel free to do so however you please.

Acknowledgments

This project builds upon the work described in Towards Monosemanticity: Decomposing Language Models With Dictionary Learning, Update on how we train SAEs, and Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet by Anthropic, and this project absolutely would not have been possible without it.