Spaces:

OpenSAE
/

README

Running

App Files Files Community

Hunter26666666 commited on about 15 hours ago

Commit

56ce7bc

verified ·

1 Parent(s): 976b173

OpenSAE: publish validated static research infrastructure and data

Browse files

Files changed (4) hide show

.gitattributes +1 -0
README.md +39 -74
assets/opensae_banner.png +3 -0
index.html +1 -0

.gitattributes CHANGED Viewed

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text

 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/opensae_banner.png filter=lfs diff=lfs merge=lfs -text

README.md CHANGED Viewed

@@ -1,74 +1,39 @@
----
-title: OpenSAE
-emoji: 🧠
-colorFrom: blue
-colorTo: purple
-sdk: static
-pinned: false
----
-<p align="center">
-  <a href="https://github.com/OpenSAE">
-    <img src="https://raw.githubusercontent.com/OpenSAE/.github/main/opensae_banner.png" alt="OpenSAE — From learned features to faithful explanations" width="100%" />
-  </a>
-</p>
-<p align="center">
-  <strong>Building open foundations for understanding, evaluating, and intervening in neural networks.</strong>
-</p>
-<p align="center">
-  <a href="https://github.com/OpenSAE"><img src="https://img.shields.io/badge/GitHub-OpenSAE-182334?style=for-the-badge&logo=github&logoColor=white" alt="OpenSAE on GitHub" /></a>
-  <a href="https://github.com/OpenSAE/SAEBase"><img src="https://img.shields.io/badge/SAEBase-Research_Infrastructure-2563EB?style=for-the-badge" alt="Explore SAEBase" /></a>
-  <a href="mailto:hma232@wisc.edu"><img src="https://img.shields.io/badge/Join_Us-Get_in_Touch-087F8C?style=for-the-badge" alt="Join OpenSAE" /></a>
-</p>
----
-## Mission
-OpenSAE advances the science of how neural networks represent information and how their internal structures shape behavior. We develop theory, methods, and open research tools to make explanations testable and interventions reproducible.
-Sparse autoencoders are a central starting point, connecting our work to broader questions in mechanistic interpretability and representation learning.
-## Research Directions
-At OpenSAE, we study how neural networks represent information and how those representations shape model behavior. Our research focuses on:
-- Understanding what learned features represent and how they are organized within neural networks.
-- Investigating how individual features and their interactions influence model behavior.
-- Studying how representations change across inputs and contexts, and what remains consistent.
-- Developing tools and rigorous evaluations to test the reliability of explanations and interventions.
-## Research & Infrastructure
-<table>
-  <tr>
-    <td width="50%" valign="top">
-      <h3>🧱 <a href="https://github.com/OpenSAE/SAEBase">SAEBase</a></h3>
-      <p>Research infrastructure for understanding, testing, and modifying SAE features in LLMs.</p>
-      <p><a href="https://github.com/OpenSAE/SAEBase">Code</a> · <a href="https://github.com/OpenSAE/SAEBase/tree/main/docs">Documentation</a></p>
-    </td>
-    <td width="50%" valign="top">
-      <h3>🧩 <a href="https://github.com/OpenSAE/SAE-Behavioral-Compositionality">Behavioral Compositionality</a></h3>
-      <p>Measures how much model behavior depends on interactions among its SAE features.</p>
-      <p><a href="https://github.com/OpenSAE/SAE-Behavioral-Compositionality">Code</a> · <a href="https://arxiv.org/abs/2610.21261">Paper</a> · <a href="https://opensae.github.io/SAE-Behavioral-Compositionality/">5-minute tutorial</a></p>
-    </td>
-  </tr>
-  <tr>
-    <td width="50%" valign="top">
-      <h3>🔀 <a href="https://github.com/OpenSAE/Redistribution-Sensitivity-Score">Representation Redistribution</a></h3>
-      <p>Detects consistent shifts in SAE feature activity when the same request is presented differently.</p>
-      <p><a href="https://github.com/OpenSAE/Redistribution-Sensitivity-Score">Code</a> · <a href="https://arxiv.org/abs/2610.21240">Paper</a> · <a href="https://opensae.github.io/Redistribution-Sensitivity-Score/">5-minute tutorial</a></p>
-    </td>
-    <td width="50%" valign="top">
-      <h3>🧭 <a href="https://github.com/OpenSAE/SAE-Input-Localization">SAE Input Localization</a></h3>
-      <p>Shows that SAE features can look simple while depending on many interacting inputs.</p>
-      <p><a href="https://github.com/OpenSAE/SAE-Input-Localization">Code</a> · <a href="https://htmlpreview.github.io/?https://github.com/OpenSAE/SAE-Input-Localization/blob/main/docs/index.html">5-minute tutorial</a></p>
-    </td>
-  </tr>
-</table>
-## Opportunities
-We share research, code, evaluation protocols, and tools to support an open, rigorous, and reproducible understanding of LLMs. If you're interested in our research and would like to join OpenSAE, please contact us at [hma232@wisc.edu](mailto:hma232@wisc.edu).

+---
+title: OpenSAE
+emoji: 🧱
+colorFrom: gray
+colorTo: blue
+sdk: static
+app_file: index.html
+short_description: Research infrastructure for understanding neural networks.
+---
+<div align="center">
+  <img src="assets/opensae_banner.png" alt="Official OpenSAE banner" width="100%" />
+  <h2>Open foundations for understanding, evaluating, and intervening in neural networks.</h2>
+  <p>OpenSAE develops open research methods and infrastructure for studying neural representations, SAE features, and their effects on model behavior.</p>
+</div>
+---
+### Research infrastructure
+**[🧱 SAEBase — Open the interactive research lab](https://huggingface.co/spaces/OpenSAE/SAEBase)**
+Capture reusable activations, intervene on SAE features, measure downstream effects, and evaluate the reliability of experiments. The published software includes 18 model–SAE configurations and 48 verified model/layer sites across models from 270M to 9B parameters.
+[GitHub](https://github.com/OpenSAE/SAEBase) · [v0.3.0 Release](https://github.com/OpenSAE/SAEBase/releases/tag/v0.3.0) · [Evaluation Data](https://huggingface.co/datasets/OpenSAE/SAEBase-Evaluation)
+### Research projects
+| Project | Research question | Explore |
+| --- | --- | --- |
+| **SAR — Behavioral Compositionality** | How do SAE features interact to shape model behavior? | [Research and tutorial](https://github.com/OpenSAE/SAE-Behavioral-Compositionality) |
+| **RSS — Representation Redistribution** | How does feature activity change across prompt presentations? | [Research and tutorial](https://github.com/OpenSAE/Redistribution-Sensitivity-Score) |
+| **SAE Input Localization** | Which controlled inputs are required to explain SAE features? | [Research and tutorial](https://github.com/OpenSAE/SAE-Input-Localization) |
+### Open science
+We share software, evidence, reproducible evaluation procedures and tutorials. Original pretrained weights remain with their publishers. Our [benchmark evidence](https://huggingface.co/datasets/OpenSAE/SAEBase-Evaluation) is freely browsable, with links to the original run records.
+**Explore:** [OpenSAE GitHub](https://github.com/OpenSAE) · [SAEBase](https://huggingface.co/spaces/OpenSAE/SAEBase) · [Contribute](https://github.com/OpenSAE/SAEBase/blob/main/CONTRIBUTING.md)
+**Contact:** hma232@wisc.edu

assets/opensae_banner.png ADDED Viewed

Git LFS Details

SHA256: d489f8310255734d36a2a3db040abff8222023fb170fa5e19881258e7c4ebd03
Pointer size: 131 Bytes
Size of remote file: 943 kB

index.html ADDED Viewed

	@@ -0,0 +1 @@

+ <!doctype html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>OpenSAE — Open Research Infrastructure</title><style>body{margin:0;background:#f5f9fa;color:#18313c;font:16px/1.65 system-ui,sans-serif}main{max-width:920px;margin:4vh auto;padding:25px}.banner{width:100%;border-radius:16px;display:block;background:#132936;max-height:270px;object-fit:cover}h1{font-size:44px;letter-spacing:-.06em;margin:22px 0 5px}p{color:#5e7883}.cards{display:grid;grid-template-columns:repeat(2,1fr);gap:12px;margin:23px 0}.card{padding:22px;border-radius:13px;background:white;border:1px solid #dfebee;color:#143342;text-decoration:none}.card strong{display:block;margin-bottom:6px}.card span{font-size:13px;color:#5d7b84}a{color:#097e79}@media(max-width:650px){.cards{grid-template-columns:1fr}h1{font-size:31px}}</style></head><body><main><img class="banner" src="assets/opensae_banner.png" alt="OpenSAE official banner"><h1>OpenSAE</h1><p>Open foundations for understanding, evaluating, and intervening in neural networks.</p><div class="cards"><a class="card" href="https://huggingface.co/spaces/OpenSAE/SAEBase" target="_blank"><strong>🧱 SAEBase Research Lab ↗</strong><span>Explore original checkpoint and feature-intervention evidence.</span></a><a class="card" href="https://huggingface.co/datasets/OpenSAE/SAEBase-Evaluation" target="_blank"><strong>▤ Evaluation Dataset ↗</strong><span>Browse 48 verified layer sites and their measurements.</span></a><a class="card" href="https://github.com/OpenSAE/SAE-Behavioral-Compositionality" target="_blank"><strong>🧩 SAR</strong><span>Behavioral interactions of SAE features.</span></a><a class="card" href="https://github.com/OpenSAE/Redistribution-Sensitivity-Score" target="_blank"><strong>🔀 RSS</strong><span>Representation redistribution across prompt changes.</span></a></div><p><a href="https://github.com/OpenSAE">GitHub organization ↗</a> · <a href="mailto:hma232@wisc.edu">Contact</a></p></main></body></html>