OpenSAE/SAEBase-Evaluation

SAEBase Evaluation

Versioned measurements for sparse autoencoder research. The original 48-site software verification tables and the later 500-prompt Research Lab experiment use different input populations and should be analyzed separately.

A unified research dataset for evaluating sparse autoencoders, tracing feature effects, and studying representation structure in language models.

SAEBase Evaluation brings model–SAE compatibility records, reconstruction and sparsity metrics, feature interventions, controlled inputs, and research measurements from SAR, RSS, and Input Localization into one Hugging Face dataset. Each record is designed to connect a numerical result to a specific model, SAE checkpoint, layer, input population, and experimental protocol.

Explore the data · SAEBase Research Lab · GitHub · v0.4.0

Research overview

SAEBase Evaluation provides reconstruction metrics, measured feature effects, and records needed to verify their source models, inputs, and evaluation protocols.

18 model–SAE configurations and 48 layer sites were covered by the v0.3.0 verification. A separate 500-prompt Research Lab cohort was used in v0.4.0. The repository provides 25 dataset subsets for these and related records.

The historical 64-group benchmark and the newer 500-group Research Lab cohort are separate experiments, not interchangeable samples. Research Lab aggregate counts and downloadable per-sample files are distinguished in the corresponding subsets.

Study populations and measurement coverage

Collection	Scale	What it helps you study
Model & SAE evaluation	18 model–SAE configurations, 48 layer sites	Sparsity, reconstruction, intervention response, checkpoint compatibility
Historical benchmark	64 held-out controlled prompt groups per site	Directly reproducible historical metrics and reliability records
Research Lab input cohort	10 sources × 50 parent groups, 200 discovery / 300 evaluation	Input sources, their pinned revisions, and separate discovery and evaluation splits
Feature and intervention records	Up to 300 feature cards at selected deeper layers and six intervention sites	Feature activity and recorded intervention effects when underlying sample records are published
SAE Additivity Residuals (SAR)	1,500 reported evaluated prompt groups across K16/K8/K4	How well individual feature effects predict their joint downstream effects
Redistribution Sensitivity Score (RSS)	Six sites, 3,000 site and parent-prompt combinations, and 138,000 reported site and input occurrences	How prompt transformations redistribute SAE feature activations
Input Localization	Two six-factor designs with 64 states each, including one grid evaluated with a real model	How feature responses depend on controlled input factors

Start in three lines

from datasets import load_dataset

metrics = load_dataset("OpenSAE/SAEBase-Evaluation", "metric_results", split="train")
print(metrics.select(range(3)).to_pandas().to_string(index=False))

Install: pip install -U datasets pandas.

The metric_results subset is the simplest entry point. Each row records one metric measured at one model–SAE layer site. Its value column is a number, metric_id identifies what was measured, n_valid gives the population denominator, and source_report_url points to the original experiment record.

Choose a subset

For reconstruction and sparsity comparisons, start with metric_results. The following entries provide the main routes into the other studies.

Research question	Starting subset	Related records
SAE reconstruction and sparsity	metric_results	historical_site_metrics
Feature additivity (SAR)	sar_cohorts	gpu_sar_per_prompt
Prompt transformations (RSS)	rss_coverage	gpu_rss_per_setting
Input localization	input_localization	controlled_inputs
Model, input, and source records	model_sae_registry	prompt_manifest, source_revisions

The cohort summary tables describe coverage. For numerical results at the level of individual inputs, use the corresponding per-record subsets when the records are published.

All 25 configurations

Evaluation and research summaries

Configuration	Contents
metric_results	Individual L0, reconstruction, explained-variance and ablation-logit metrics
historical_site_metrics	One wide metric row per verified historical site
historical_full_metrics	Detailed historical metric rows, including the identity intervention
historical_model_aggregates	Per-model/SAE summaries across selected historical layers
historical_acceptance_probes	Individual real-checkpoint probe results
research_lab_summary	Measurements reported by the Research Lab executor

SAR and RSS

Configuration	Contents
sar_cohorts	K16/K8/K4 cohort counts and coalition configurations
gpu_sar_per_prompt	Measured prompt-level SAR values and response/checksum references
rss_coverage	Distinct input coverage, selection/held-out counts and status
gpu_rss_per_setting	RSS measurements by setting and persistence measurements on held-out variants

Inputs and localization

Configuration	Contents
prompt_manifest	500 frozen source-group identifiers and their splits, without third-party prompt text
source_revisions	Pinned upstream dataset revisions and source digests
source_split_counts	Input counts by source and discovery/evaluation split
controlled_inputs	Complete project-authored six-factor input texts
controlled_designs	Design identities and 64-state experiment definitions
input_localization	Feature variation and scope cost measured on the complete 64-state input grid
gpu_localization_records	Controlled-grid metric and numerical diagnostic records

Model and checkpoint records

Configuration	Contents
verified_sites	48 model/SAE layer sites verified in v0.3.0
model_sae_registry	Model/SAE layer and hook registry
model_sae_profiles	One row per model–SAE configuration
checkpoint_identities	Recorded checkpoint digests, hooks and runtime identity
gpu_site_capture_ledger	Per-site capture status and recorded model/checkpoint hashes
gpu_new_checkpoint_identities	Verified new-run identities distinct from historical ones

Feature and intervention records

Configuration	Contents
gpu_feature_cards	Evidence-linked feature activity cards
gpu_intervention_pairs	Model outputs recorded before and after interventions

Supporting CSV files and source records

Some supporting JSON and CSV files are also browsable from the Files tab. The available downloadable research resources are listed below when detected in the repository.

File	What it contains
prompt_manifest.csv	Frozen input hashes and group IDs
sources.csv	Original datasets and pinned revisions
sites.csv	Layer-site registry
controlled_inputs.csv	Project-authored 64-state designs
sar_cohorts.csv	SAR cohort summary
rss_coverage.csv	RSS coverage summary
localization.csv	Localization summary
sar_prompt_metrics.csv	Measured SAR prompt records
rss_group_settings.csv	Measured RSS group/setting records
feature_cards.csv	Feature-level evidence cards
intervention_pairs.csv	Observed intervention pairs
all_metric_results.csv	All source-pinned historical metric rows
acceptance_probes.csv	Checkpoint acceptance probes

A row is not always an independent observation. The statistical unit is the parent prompt group for natural-input studies and the complete 64-state grid for each controlled Localize design.

Column definitions

Column	Meaning
model_id, sae_id, layer	Which pretrained language model and SAE dictionary were measured, and at which layer
site_id, hook_name	Exact model/SAE intervention location; check the schema before joining tables
metric_id, condition, value	Metric name, experimental edit/condition (if any), and numerical result
n_valid, status	Measurement denominator and whether the result is available or defined
sample_id, prompt_group_id	Input identity and the independent group to which variants belong
usage_split	Discovery or evaluation assignment; transformed versions inherit their parent group
source_commit, source_report_url, *_sha256	Code and data provenance and links for verification
protocol, k, M, n_fit, n_held	Scientific method and its coalition/sample configuration for SAR
selection_variant_ids, held_out_variant_ids	RSS variants used for selection and independent evaluation when raw records are present

Practical examples

1. Find the most reconstructable layer in one SAE

from datasets import load_dataset

D = load_dataset("OpenSAE/SAEBase-Evaluation", "metric_results", split="train").to_pandas()
model = "gemma270m"
sae = "gemma270m_scope_resid"

layers = D[(D.model_id == model) & (D.sae_id == sae)]
metrics = layers.pivot_table(index="layer", columns="metric_id", values="value", aggfunc="first")
print(metrics[["core.l0", "core.reconstruction_mse", "core.explained_variance"]].sort_index())

A smaller reconstruction MSE indicates more accurate reconstruction under the activation scale used in that experiment. Explained variance uses a variance-based denominator. Compare layers within a compatible protocol before comparing results from different models or hooks. See the SAEBase repository for benchmark examples and scripts.

2. Examine the effect of feature ablation

from datasets import load_dataset

D = load_dataset("OpenSAE/SAEBase-Evaluation", "metric_results", split="train").to_pandas()
ablations = D[(D.metric_id == "intervention.logit_delta_l2") & (D.condition == "feature-ablation")]
print(ablations[["model_id", "sae_id", "layer", "value", "n_valid"]].head(10))

Here, value is an L2 change in the model's output logits caused by the recorded edit. It indicates a downstream effect, not by itself a proof that the selected feature has a unique human-readable meaning.

3. Retrieve the 500-prompt source manifest

If your version of the Dataset includes the prompt_manifest subset:

from datasets import load_dataset

prompts = load_dataset("OpenSAE/SAEBase-Evaluation", "prompt_manifest", split="train").to_pandas()
print(prompts.groupby(["dataset_id", "usage_split"]).size().unstack(fill_value=0))

The manifest contains identifiers, grouping, upstream revisions and text digests, not necessarily the original copyrighted prompt text. This makes it useful for reproducing dataset selection, checking leakage and comparing runs on the same inputs.

4. Work with SAR, RSS and Localize results

from datasets import load_dataset

repo = "OpenSAE/SAEBase-Evaluation"

# Study populations and the selected coalition configurations.
sar = load_dataset(repo, "sar_cohorts", split="train").to_pandas()
print(sar[["k", "m", "fit", "held", "successful_groups"]])

# Counts in these tables are measurements of coverage, not per-prompt effect sizes.
rss = load_dataset(repo, "rss_coverage", split="train").to_pandas()
print(rss[["parent_site_groups", "unique_site_input_occurrences"]])

loc = load_dataset(repo, "input_localization", split="train").to_pandas()
print(loc[["site_id", "states", "observed_support_features", "budget_at_90pct"]])

SAR: A coalition is a selected subset of SAE features. The original SAR estimator fits and evaluates an additive approximation to measured downstream responses. In the separate Research Lab experiment, an edited activation is computed as edited_h = h + decode(z_edited) - decode(z_original). This retains the part of the original activation not reconstructed by the SAE. The K16, K8 and K4 cohorts are analyzed separately. The sar_cohorts table reports group counts, not the numerical SAR residual for each prompt. See the SAR paper for the original method. When provided, gpu_sar_per_prompt contains the actual per-prompt estimates and checksums of their underlying response vectors.

RSS: This analysis measures changes in SAE activations under prompt transformations. rss_coverage counts how many site and input combinations were evaluated. These counts do not represent independent observations or an average RSS value. When provided, gpu_rss_per_setting reports selected feature sets, summaries of the direction of activation changes, and persistence on held-out variants. The proposed transformations have not all been verified to preserve the meaning of the original request. See the RSS paper for the original method.

Localization: A six-factor experiment enumerates all (2^6 = 64) input states, then examines how much of the resulting feature variation can be represented by limited subsets of input factors. controlled_inputs contains project-authored states; input_localization is a separate real model result. The recorded budget_at_90pct follows the method's exact scope-cost definition; do not interpret it as simply a count of nonzero SAE dictionary features.

For richer per-record analysis, use these subsets if shown in the available-configurations table: historical_full_metrics, historical_acceptance_probes, gpu_sar_per_prompt, gpu_rss_per_setting, gpu_feature_cards, gpu_intervention_pairs, and gpu_localization_records. Aggregate summaries must not be treated as substitutes for unuploaded per-prompt measurements.

Research Lab experiments (v0.4.0, separate cohort)

SAEBase also reports a separate, frozen 500-prompt research cohort for feature activations, interventions, and additional analyses. Research Lab offers interactive browsing of recorded measurements, and the GitHub evidence summary documents the experiments and their scope.

This is not the same population as the 64-group-per-site v0.3.0 verification tables. Only files actually published in this dataset repository should be treated as downloadable dataset assets; a recorded GPU experiment does not imply its complete private raw archive is publicly released. Keep the cohorts and their source manifests separate.

How the dataset is organized

The dataset organizes these records while keeping experiments with different protocols separate.

SAEBase-Evaluation (one dataset repository)
 ├── Historical evaluation      model/SAE/layer × metric, linked to source reports
 ├── Model and site registry    checkpoint identities, hooks, layers
 ├── Input manifests           source revisions, groups and splits
 ├── Controlled designs        64-state input grids created for this project
 ├── Feature & interventions   recorded codes and edits, when published
 ├── SAR / RSS                  group and setting metrics, when published
 └── Localization              feature responses to controlled input factors

Use (model_id, sae_id, layer) to connect historical metrics to the newer registry. In particular, some older site_id values abbreviate the model/layer, while the Research Lab site IDs also include the SAE ID; do not assume they are directly equal as strings.

The historical v0.3.0 results come from the verified checkpoint matrix and original raw run archives. The newer Research Lab results are documented in the v0.4.0 experiment report with the relevant source releases, model identities and measurement protocols.

Use it for your research

This dataset is designed for three practical workflows:

• Compare reconstruction, sparsity and measured feature effects across layers or SAE checkpoints while preserving protocol identities.
• Audit and reproduce an existing result by tracing a table row back to its source report, source code revision and checkpoint digest.
• Develop new interpretability analyses using SAR, RSS, controlled inputs and feature-level records, when the needed per-sample arrays are available in the dataset or in referenced experiment archives.

For live exploration or to try the related code, use SAEBase Research Lab. This dataset publishes research records and inputs, not pretrained SAE weight binaries or an online inference service.

Citation

If you use these records, cite the SAEBase software version and original model/SAE sources. For results based on SAR or RSS, also cite the corresponding paper. SAEBase repository · SAEBase v0.4.0.

• SAR: Haotian Ma, Ziye Mei, and Frederic Sala. Behavioral Compositionality of Sparse Autoencoder Representations. 2026. Hugging Face paper · arXiv:2610.21261 · GitHub.
• RSS: Ziye Mei and Haotian Ma. Representation Redistribution in Sparse Autoencoders. 2026. Hugging Face paper · arXiv:2610.21240 · GitHub.

BibTeX entries for both papers are available in CITATION.bib.

---

_{Maintained by OpenSAE. Results retain their original experimental context and upstream source identities.}