Skip to content

SAEBase v0.4.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 20 Oct 16:53
· 351 commits to main since this release
v0.4.0
c0a59ba
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

SAEBase v0.4.0 — Research Lab Preview

New in v0.4.0: SAEBase graduates from a Python-centered SAE evaluation toolkit to a local, browsable Research Lab workflow, with a new measured 500-prompt GPU cohort reported separately from historical 64-group engineering acceptance.

What you can use

  • Research Lab browser: results search, feature index, controlled intervention preflight, SAR/RSS/Localize analysis pages, dataset ledger and immutable report API. Run with the CPU Docker image in the repository.
  • New source cohort: ten pinned public dataset sources × 50 original groups each; 200 discovery/300 evaluation. Source data stays private until content/rights approval.
  • Real GPU execution: code and manifest conventions for 48 registered SAE sites on the new corpus, six deep-site interventions and feature card extraction.
  • Original-estimator integrations: SAR and RSS use their original paper estimators; a separately named decoder-mediated SAR intervention mode permits direct evaluation of measured coalition responses where the affine decoder precondition is not established.
  • Controlled Localize: exact, complete six-factor input-grid handling and a real model support-only numerical curve.
  • Engineering: stable source/checkpoint data contracts, re-entrant work units, source-bound regression testing, paginated feature store, private queue state machine and read-only CPU deployment.

New measured progress, as reported by GPU executor

Measurement Reported result
New prompt population 10 × 50 = 500 frozen source groups
Base activation/SAE capture 48 sites × 500 groups
Deep-site statistical feature cards 300 across six sites
SAR 1,485 K16 + 14 K8 + 1 K4 successful groups
RSS 138,000 logical unique site-input occurrences; six sites; 500 groups/site
Localize 64-state real grid; 55 observed-support columns; recorded budget 132
Intervention Six sites with measured clean/edited logits

Interpretation and release boundary: These are reported measurements from a private GPU experiment; the full experiment archive has not been independently re-downloaded as part of this software release. RSS transformations are not verified meaning-preserving. Decoder-mediated SAR differs from SAR under an affine decoder assumption. The permutation of feature order is not an orthogonal rotation control. Full steered text generation, public GPU serving and HF Space deployment remain outside this release.

See the detailed Research Lab report and release acceptance status.

Install and run

git clone --branch v0.4.0 https://github.com/OpenSAE/SAEBase.git
cd SAEBase
python -m pip install -e .
python examples/quickstart/public_api_offline.py

Read-only Research Lab on CPU:

docker build -f apps/research-lab/Dockerfile -t opensae-lab:cpu .
docker run --rm -p 7860:7860 opensae-lab:cpu

Open http://127.0.0.1:7860. The Docker image contains historical published evidence, not the new private GPU archive.

The versioned GitHub software Release, when published, contains the verified wheel and source archive plus source/asset integrity evidence. No PyPI upload is implied.

Compatibility and scientific provenance

  • Python 3.11–3.12, existing nine CLI entrypoints and SAEBase imports are retained.
  • The default SAR interface keeps require_affine_decode=True; the decoder-mediated experiment requests the alternative protocol explicitly.
  • No vLLM/KV-cache GPU parity is claimed.
  • Third-party checkpoint weights, non-public prompts and GPU raw data are not redistributed.
  • Previously published v0.2.0/v0.3.0 releases and Stage D evidence are unchanged.

Links: Research Lab · Model matrix · Work Plan 05 · Changelog.

Assets 6
Loading