README.md

---
title: OpenSAE
emoji: 🧱
colorFrom: green
colorTo: gray
sdk: static
app_file: index.html
short_description: Open research for testable explanations of neural networks.
---

# From features to testable explanations.

**Open research tools for understanding how neural networks represent information and how their features shape behavior.**

[**Explore SAEBase Research Lab ↗**](https://huggingface.co/spaces/OpenSAE/SAEBase)   ·   [**See our research ↗**](https://huggingface.co/spaces/OpenSAE/README)

<a href="https://huggingface.co/spaces/OpenSAE/README"><img src="https://huggingface.co/spaces/OpenSAE/README/resolve/main/assets/opensae_banner.png" alt="OpenSAE" width="240"></a>

## SAEBase

**Capture. Intervene. Evaluate.**

One reusable workflow for studying SAE features and the effects of interventions.

**[Open the interactive lab ↗](https://huggingface.co/spaces/OpenSAE/SAEBase)** &nbsp; · &nbsp; [Source code](https://github.com/OpenSAE/SAEBase) &nbsp; · &nbsp; [Evaluation data](https://huggingface.co/datasets/OpenSAE/SAEBase-Evaluation)

## Research

**[SAR ↗](https://github.com/OpenSAE/SAE-Behavioral-Compositionality)** — How do features work together to shape behavior?

**[RSS ↗](https://github.com/OpenSAE/Redistribution-Sensitivity-Score)** — What changes inside a model when the prompt changes?

**[Input Localization ↗](https://github.com/OpenSAE/SAE-Input-Localization)** — Which inputs are responsible for feature activations?

---

[**GitHub organization ↗**](https://github.com/OpenSAE) &nbsp; · &nbsp; [Contribute](https://github.com/OpenSAE/SAEBase/blob/main/CONTRIBUTING.md) &nbsp; · &nbsp; [Contact](mailto:hma232@wisc.edu)