File size: 2,410 Bytes
63290ad 9d22f2f 63290ad 9d22f2f 63290ad 9d22f2f 2e6a0cf 9d22f2f 2e6a0cf 9d22f2f 2e6a0cf 9d22f2f 2e6a0cf 9d22f2f 2e6a0cf 9d22f2f 63290ad 9d22f2f 63290ad 9d22f2f dadd096 63290ad d74aad2 2e6a0cf 9d22f2f 2e6a0cf 9d22f2f 63290ad 9d22f2f 63290ad 2e6a0cf 9d22f2f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 | ---
title: OpenSAE
emoji: 🧱
colorFrom: gray
colorTo: blue
sdk: static
app_file: index.html
short_description: Open research in mechanistic interpretability.
---
<div align="center">
<p style="font-size:16px;letter-spacing:0.14em;font-weight:700;color:#65748b;">OPENSAE</p>
<h1 style="font-size:42px;line-height:1.13;letter-spacing:-0.045em;margin:14px 0 18px;">From learned features<br/>to testable explanations.</h1>
<p style="font-size:18px;line-height:1.6;max-width:620px;margin:0 auto 26px;">Open research tools for understanding how neural network features represent information and shape model behavior.</p>
<p>
<a href="https://huggingface.co/spaces/OpenSAE/SAEBase"><strong>Explore Research Lab ↗</strong></a>
<a href="https://github.com/OpenSAE/SAEBase">Source Code ↗</a>
</p>
</div>
---
## SAEBase Research Lab
| 18 model–SAE configurations | 48 layer sites | 500 research prompts |
| :---: | :---: | :---: |
| **1,500 SAR groups** | **138k RSS observations** | **64 controlled states** |
A single place to explore representations, inspect interventions and compare reproducible research results.
**[Open the interactive Lab](https://huggingface.co/spaces/OpenSAE/SAEBase)** · **[v0.4.0 Release](https://github.com/OpenSAE/SAEBase/releases/tag/v0.4.0)**
## Explore the datasets
| Dataset | Contents |
| :--- | :--- |
| [SAEBase Evaluation](https://huggingface.co/datasets/OpenSAE/SAEBase-Evaluation) | 192 historical metrics, 48 sites, 500 prompt identities, SAR/RSS/Localization tables, and 128 controlled inputs |
## Research
**[SAR](https://github.com/OpenSAE/SAE-Behavioral-Compositionality)** · Feature interactions and downstream behavior
**[RSS](https://github.com/OpenSAE/Redistribution-Sensitivity-Score)** · Representation redistribution under prompt changes
**[Input Localization](https://github.com/OpenSAE/SAE-Input-Localization)** · Which input factors explain feature activations
**[SAEBase](https://github.com/OpenSAE/SAEBase)** · Infrastructure for reusable SAE evaluation
---
<div align="center">
<p><strong>OpenSAE · Open research for mechanistic interpretability</strong></p>
<p><a href="https://github.com/OpenSAE">GitHub</a> · <a href="https://huggingface.co/spaces/OpenSAE/sar-tutorial">SAR in Five Minutes</a> · <a href="https://huggingface.co/OpenSAE/collections">Collections</a></p>
</div>
|