SAEBase v0.3.0

SAEBase v0.3.0 Release Notes

SAEBase v0.3.0 extends the original v0.2.0 research toolkit with broader real
model/SAE compatibility, reproducible population engineering checks, a versioned
external extension API, source-bound benchmarking and expanded onboarding.

Highlights

• Model and SAE coverage: 18 selected model/SAE profiles and 48 model/SAE
  layer sites (270M–9B) with recorded real-pair acceptance, 64-group engineering
  benchmarks and reliability checks.
• Evidence: The exact original Stage D raw archives and compact experiment
  records are available in separate checksum-verifiable GitHub Releases.
  Original Stage D scientific results and provenance are preserved.
• Third-party extensions: Explicitly register external models, SAEs and
  evaluation metrics through source-aware Python entry points, with strict
  layer, hook, dimension, dtype and normalization contracts.
• Research workflow: Reusable activation stores, feature interventions,
  controlled input grids, source-pinned SAR/RSS/Input Localization integrations
  and native SAEBench result imports.
• Developer experience: A persistent offline first experiment, bilingual
  README, contributor guides, package tests on Python 3.11/3.12, and a
  source-bound software release verification workflow.

Install

Download the Python wheel and SHA256 manifest from the v0.3.0 GitHub Release
once published, or install from the tagged source checkout.

git clone --branch v0.3.0 https://github.com/OpenSAE/SAEBase.git
cd SAEBase
python -m pip install -e .
python examples/19_first_experiment_offline.py --output runs/first-experiment

For actual pretrained model/SAE runs, install .[analysis] and follow
Getting Started. Real weights are downloaded from
their original publishers when needed, not redistributed with SAEBase.

Verification and evidence

• Verified model/SAE matrix
• Stage D evidence documentation
• Original Stage D raw release
• Stage D compact release
• P1/P2 independent exploratory archive

These records are engineering interoperability and benchmark-protocol
acceptances, not evidence of general interpretability quality or broad
natural-language generalization. The P1 probe covers one external plugin
checkpoint and hook, not all third-party extension packages.

Compatibility

The v0.2.0 source, tag and published assets remain unchanged. Existing CLI
entry points and import shims are retained. The external extension API is
new in the v0.3.0 source. Third-party checkpoint and dataset licenses remain
with their respective publishers.

Release only after six required CI jobs pass at the exact candidate main
commit and the manual software release verifier succeeds.

Stage D recovery 03 experiments (P1 and P2)

Archival package for the work-plan-03 GPU experiments that are not already in the historical Stage D raw release.

• P1 external-extension probe for Qwen3-1.7B Base L7
• P2 exploratory 64-group runs for Qwen3-1.7B Base L7 and Qwen3-8B Base L9
• Native 48/48 Stage D audit records

This does not replace https://github.com/OpenSAE/SAEBase/releases/tag/stage-d-original-raw-20261016 and does not change v0.2.0 or stage-d-evidence-20261016.

No pretrained weights or credentials are included. The independent download receipt is uploaded after this archive is sealed.

Stage D original raw archives 20261016

Stage D original raw evidence (2026-10-16)

Immutable original GPU archives for the Stage D population M3/M4 runs.
This release is separate from the Python package v0.2.0 and from the compact
evidence release stage-d-evidence-20261016.

• Source repository: OpenSAE/SAEBase
• Experiment evidence commit: 568e7ba5df1f9e27ed32e42299d588a497fc9f6e
• Packaging / verifier commit: a59e013447445d213f8ba3d86ca9485cdab044d9
• 18 original family archives, 725190 bytes, status=byte_hash_verified
• Combined bundle: stage-d-evidence-20261016-original-raw.tar.gz
• Native M3/M4 audit of the extracted runs: 48/48 native_m3_m4_audited
• Contents are benchmark run directories, compact reports, and token ledgers.
  No pretrained model or SAE weights and no credentials.

Verify the combined bundle:

sha256sum -c stage-d-evidence-20261016-original-raw.tar.gz.sha256
python scripts/stage_d_archive.py --verify-raw-bundle stage-d-evidence-20261016-original-raw.tar.gz

The individual *_population_*.tar.gz files are the unmodified originals named
by conf/coverage/stage_d_evidence/final/raw_run_archive_index.json.
Check them with sha256sum -c ORIGINALS.sha256.

Stage D compact verification evidence (48 sites)

Stage D Evidence Archive

The Stage D compact evidence release provides a versioned snapshot of original committed evidence records, distinct from the SAEBase Python package release.

Contents

• The frozen 18-profile, 48-layer Stage D catalog, final matrix and per-layer status ledger.
• 48 M3/M4 compact summaries, 48 selection and 48 evaluation token ledgers, and the Stage D pair-acceptance JSONs. Three earlier CPU pair acceptances retain their original Stage B ledger rather than fabricated new GPU records.
• All 200 checked-in Stage D evidence files, including the manifest covering 199 of them (excluding the checksum manifest itself).
• Original Stage B/C ledgers and the frozen Stage D evaluation corpus.
• A file-by-file SHA256 bundle manifest, compressed archive checksum, exact Git source revision and a self-contained Python verifier.

The reports record 48/48 real model-SAE pair and 48/48 population benchmark/reliability engineering acceptances. The evaluation uses 64 separate groups and recurring controlled prompt templates. These results demonstrate the reported engineering protocol, not general SAE quality or interpretability.

Scope of the archive

The compact release does not contain original GPU run caches or pretrained weights. The original raw-run archive index names 18 GPU archive files totaling 725,190 compressed bytes but stores a machine-local path, not a public download link. Original archives were not found among the repository files, existing Release assets or relevant GitHub Actions artifacts inspected during this work.

The compact evidence is independently hash-verifiable, but it cannot on its own support recomputing all metrics from the unavailable raw M3/M4 run directories. The original README.json value of 17 M3/M4 summaries is an earlier in-progress snapshot; the committed final matrix and 48 per-site reports are the final source. These historical records have not been rewritten.

Download and verify

Download all four assets from the GitHub evidence release or run:

gh release download stage-d-evidence-20261016 -R OpenSAE/SAEBase --dir stage-d-download
cd stage-d-download
sha256sum -c SHA256SUMS
python verify-stage-d.py --verify-bundle stage-d-evidence-20261016-compact.tar.gz

The first command checks download integrity; the second safely extracts into a temporary location, checks each file against its SHA256 manifest, then validates the original 199 Stage D hashes, catalog and status ledger and all 48 per-site reports without model weights.

A source checkout can be verified or repackaged entirely offline:

python scripts/stage_d_archive.py --check
python scripts/stage_d_archive.py --build --output-dir dist/stage-d-evidence
python scripts/stage_d_archive.py --verify-bundle dist/stage-d-evidence/stage-d-evidence-20261016-compact.tar.gz
(cd dist/stage-d-evidence && sha256sum -c SHA256SUMS)

The bundle's source-commit SHA records the packaging revision, distinct from the historical GPU experiment commits contained in the Stage D JSON.

Recover original raw M3/M4 archives

Retrieve the original bytes from the GPU machine or an existing backup, then run the following against the original archive directory:

python scripts/stage_d_archive.py --verify-raw /workspace/stage_d_runs/archives
python scripts/stage_d_archive.py --build-raw /workspace/stage_d_runs/archives --output-dir dist/stage-d-raw
python scripts/stage_d_archive.py --verify-raw-bundle dist/stage-d-raw/stage-d-evidence-20261016-original-raw.tar.gz
(cd dist/stage-d-raw && sha256sum -c stage-d-evidence-20261016-original-raw.tar.gz.sha256)

Verification refuses absent files, incorrect sizes or SHA256 digests, and unsafe TAR members. Preserve the originals in a separate immutable public release only after these checks pass. Restore their raw run directories and apply the existing scripts/phase_d_verify_evidence.py for a subsequent audit against native metrics, manifests and compact summaries. Hash checking alone is not re-execution or independent numerical verification.

Original model/SAE weights remain subject to upstream access and license restrictions and are not included.

SAEBase v0.2.0 — M0–M5

SAEBase v0.2.0

The first P0 release of OpenSAE SAEBase. M0–M5 deliver reference checkpoint
acceptance, activation caching, intervention evaluation, multi-metric
benchmarking, statistical reliability auditing and a real checkpoint
release matrix.

What's included

• Python 3.11 and 3.12 distribution with the original SAEBase API and
  saebase-benchmark, saebase-audit, saebase-ground-truth and
  saebase-release-matrix command-line interfaces.
• Two verified public language-model families: Pythia-70M deduped and
  GPT-2 Small, two real SAE checkpoints per family at the same hook.
• Six fixed prompt groups per model, identity and reconstruction conditions,
  verified activation/code sharing, original metric measurements, grouped
  confidence intervals, and hash-linked portable reports.
• Checkpoint identity SHA-256, execution time and RSS metadata, synthetic
  failure-injection resume tests and a machine-verifiable release gate.
• No global SAE-quality score, model/SAE weight binaries, or third-party
  evaluation datasets bundled into the package.

Install

Install the wheel attached to this GitHub Release with Python 3.11 or 3.12.
Optional analysis integrations require their separate dependencies. Real
checkpoint reproduction downloads upstream public weights on demand.

python -m pip install opensae_saebase-0.2.0-py3-none-any.whl
saebase-release-matrix --help

The package is under the stated SAEBase Apache-2.0 software license.
Each external model/SAE remains subject to its own upstream license.
Some SAE checkpoint redistribution rights are not independently verified,
so weights are never included here. This is not a license certification.

The four-checkpoint, six-prompt real run validates the engineering pipeline.
It does not establish universal model behavior, population confidence limits,
interpretability quality or a comprehensive benchmark ranking. The optional
extended SAE widths, gated Gemma models, and GPU generation backends are not
covered by this minimum release matrix. No PyPI deployment is implied.

For reproducible protocols, identifiers, provenance and release validation
steps, see M5 release matrix.