▣
Activation store
Capture once, reuse across SAE checkpoints and experiments with identity-aware caching.
RELEASE v0.3.0 · OPEN RESEARCH INFRASTRUCTURE
Study the features.
Measure the behavior.
SAEBase connects activation capture, feature interventions, quantitative benchmarks, and reliability analysis in one reproducible research workflow.
MODEL–SAE CONFIGURATIONS
18across 270M–9B parameter modelsVERIFIED MODEL/LAYER SITES
48real checkpoint compatibilityBENCHMARK + RELIABILITY
48 / 48recorded evaluations and auditsPUBLIC SOFTWARE
v0.3.0GitHub Release · Apache-2.0CAPABILITIES
One workflow, from representations to evidence
⌁
Feature interventions
Ablation, scaling, clamping, coalition edits and downstream behavioral measurements.
▥
Benchmarking
Reconstruction, sparsity and intervention measurements with frozen experiment configurations.
◎
Reliability
Group-aware statistical auditing, source provenance, persistent reports and repeatability checks.
Traceable evaluation. Browse each checkpoint, inspect its recorded measurements, and open the original source report. Results follow a controlled evaluation protocol; they are not a cross-model quality ranking.
VALIDATED CHECKPOINTS
Verified model & SAE coverage
Browse 48 real-weight model/layer sites across 18 SAE configurations. Select any row to inspect the exact hook, metrics, and original verification report.
SELECTED SITES
48all passed compatibility checksCONFIGURATIONS
18TopK · JumpReLU · variantsHELD-OUT EVALUATION
64controlled prompt groups per siteCheckpoint browser
Filter by checkpoint, architecture or model family.
48 records
| MODEL | SAE ARCHITECTURE | LAYER | COMPATIBILITY | RESULT |
|---|
Select a checkpoint row to inspect its measurements and source report.
These values refer to selected model/SAE hooks, not every possible checkpoint. The original source contains weights provenance, report identities, and independent reliability auditing.
ORIGINAL EVALUATION RECORDS
Explore recorded measurements
Inspect precomputed numeric results from genuine checkpoint runs. No model execution happens in this browser.
Within-checkpoint layer profile
Choose a model and a metric; comparisons are confined to its recorded layers.
Values are recorded engineering measurements on 64 held-out controlled prompt groups per site. Reconstruction scales and readout units may differ across checkpoints; do not compare model families as a quality ranking.
INSPECT THE SOURCE
Follow every measurement back to the experiment
RESEARCH WORKFLOW
From activation to reliable evidence
A single framework for collecting SAE representations, measuring interventions and reporting results.
01 · INPUTS
→Model & SAE
Identify source checkpoints, exact layers and activation hooks.
02 · CACHE
→Activation store
Capture and reuse model activations and sparse feature codes.
03 · EDIT
→Feature interventions
Modify selected features and measure downstream effects.
04 · AUDIT
Benchmarks & reliability
Produce versioned metrics and statistically audited reports.
RESEARCH INTEGRATIONS
Tools built on the same foundations
FEATURE INTERACTIONS
SAR
Measure non-additive downstream effects between SAE features.
Explore research ↗PROMPT TRANSFORMATIONSRSS
Study consistent feature redistribution across input presentations.
Explore research ↗INPUT DEPENDENCIESInput Localization
Study the inputs needed to explain learned feature activations.
Explore research ↗Research estimators remain in their own repositories; SAEBase supplies common capture, intervention, evaluation and provenance infrastructure.
5-MINUTE TOUR
Your first reproducible SAE experiment
Follow the key stages of a SAEBase run. This walkthrough explains the real software workflow using examples and links; no model runs in the browser.
THE WORKFLOW
6 short steps · Python 3.11–3.12 · start without GPU
STEP 01 / 06
Start with a known environment
ACTIVATION → SAE → INTERVENTION → MEASUREMENT
1 of 6
Use the linked GitHub tutorial to execute the experiment locally. The browser tour explains which files are generated and how they are verified; it does not substitute synthetic values for model evidence.
FOR DEVELOPERS & RESEARCHERS
Build with SAEBase
Begin with an offline reproducible experiment, then bring your own model, SAE or benchmark metric.
Five-minute local quickstart
Python 3.11–3.12 · no pretrained checkpoint or GPU required for this first example.
git clone https://github.com/OpenSAE/SAEBase.git
cd SAEBase
python -m pip install -e .
python examples/quickstart/first_experiment.py --output runs/first-experimentThe quickstart follows the current GitHub main branch and uses a deterministic toy model. The dashboard measurements remain pinned to the independently published v0.3.0 evaluation records.