SAEBase v0.5.0

SAEBase v0.5.0 — Public Research Beta

SAEBase v0.5.0 introduces a sample-first browser Workspace, local GPU-backed feature research workflows and additional source-level infrastructure for mechanistic interpretability.

Try the OpenSAE Workspace · Python API · Source

What you can do

• Explore: inspect token activations and sparse autoencoder features with an approved Gemma example, or capture new features using an authorized local GPU worker.
• Intervene: choose a feature and editing operation, inspect original recorded measurements, and run authenticated teacher-forced edits on a compatible local GPU.
• Generate: compare baseline and edited continuations. The local worker supports 16–32 new greedy tokens for identity, ablation and scaling with source-identified checkpoints.
• Circuits: browse a recorded native Gemma-2-2B feature attribution graph (155 nodes and 748 displayed edges) with node selection and edge filtering.
• Evaluate: inspect source-linked native SAEBench Core and Sparse Probing metrics. The existing example is a single measured pilot, not a benchmark leaderboard.
• Train: explore original 50-step Sparsify TopK checkpoint and reconstruction evidence. Alternative configurations in the public browser are explicitly labeled Demo unless rerun with native training.
• Interpret: browse held-out diagnostics; an existing candidate with score 0.50 is unapproved. No unreliable natural-language feature explanation is claimed as verified.

The browser provides example inputs and functional controls even when no GPU service is connected. Measured / Demo / computed-on-worker states are distinguished.

Python package and local API

Requirements: Python 3.11 or 3.12.

git clone --branch v0.5.0 https://github.com/OpenSAE/SAEBase.git
cd SAEBase
python -m pip install -e ".[analysis]"
python examples/quickstart/public_api_offline.py

Read-only Research Lab:

python -m pip install -r apps/research-lab/requirements.txt
cd apps/research-lab
uvicorn backend.api:app --host 127.0.0.1 --port 8000

Browse http://127.0.0.1:8000/api/docs. See six local workflows for real GPU capture, steering, SAEBench, circuit tracing, training and auto-interpretation commands. This release retains the existing nine SAEBase CLI entrypoints.

Distribution and service boundaries

The GitHub software release contains a tested wheel, source archive, integrity manifest and SHA256 checksums only. There is no implied PyPI publication, bundled checkpoint/model weights or raw private GPU data. Other rights-reviewed compact records remain in source or the public website, separate from the software package.

GPU compute is not universally available: live new-prompt inference requires a separately running compatible worker, explicit operator authentication and suitable model/SAE weights. The test-only Cloudflare Quick Tunnel is transient and not a stable production API address. The static website and existing recorded results remain accessible without it.

The WP12 source-level 48/50 core-operation assessment is not an independently released 50-function platform or a guarantee of quality across all upstream models. Earlier SAR/RSS scientific caveats remain unchanged. This is a public software Beta release, not an independently validated raw-data release.

Security

The public website never embeds an operator token. Rotate credentials exposed during prototyping before offering GPU access; use strong distinct operator/worker secrets, private storage, authenticated routes, quotas, and a stable secured reverse proxy before wider public compute. CORS is a browser policy, not user authentication.

See changelog, release process and WP17 feature handoff.