| title: SAEBase Research Lab | |
| emoji: 🧱 | |
| colorFrom: gray | |
| colorTo: blue | |
| sdk: static | |
| app_file: index.html | |
| short_description: Explore verified SAE checkpoints, interventions and results. | |
| tags: | |
| - sparse-autoencoders | |
| - mechanistic-interpretability | |
| - research-tools | |
| - reproducibility | |
| - evaluation | |
| pinned: true | |
| # SAEBase Research Lab | |
| Explore SAEBase workflows, real-checkpoint compatibility results and source-linked evaluation data in an interactive research dashboard. | |
| Start with the **5-minute tutorial**, browse 18 model–SAE configurations and 48 verified layer sites, then inspect four recorded metrics and original per-layer reports. The dashboard displays precomputed results rather than running an LLM in the browser. | |
| All measurements were obtained from the verified source at [`d3fd1cb1`](https://github.com/OpenSAE/SAEBase/tree/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279). Metric values may not be directly comparable across models, hooks, and evaluation protocols. Benchmark prompts use 64 held-out groups of recurring controlled templates. | |
| **Source:** https://github.com/OpenSAE/SAEBase | |
| **Release:** https://github.com/OpenSAE/SAEBase/releases/tag/v0.3.0 | |
| **Evaluation files:** https://github.com/OpenSAE/SAEBase/blob/main/docs/verified-model-matrix.md | |
| The Space hosts public software demonstrations and source-linked evaluation records. Pretrained models and SAE weights remain with their original publishers. | |