---
pretty_name: SAEBase Model and SAE Evaluation Records
tags:
  - sparse-autoencoder
  - mechanistic-interpretability
  - evaluation
  - reproducibility
  - benchmark
language:
  - en
configs:
  - config_name: verified_sites
    data_files:
      - split: train
        path: data/verified_sites.csv
  - config_name: metric_results
    data_files:
      - split: train
        path: data/metric_results.csv
---

# SAEBase Evaluation Records

**48 verified model/SAE layer sites · 18 model–SAE configurations · SAEBase v0.3.0**

This dataset collects original-source, machine-readable verification records for the SAEBase software evaluation workflow. Every row links to its original per-site report in the [SAEBase repository](https://github.com/OpenSAE/SAEBase). The published source commit is [`d3fd1cb1`](https://github.com/OpenSAE/SAEBase/tree/d3fd1cb1a453118611f3c7d8b44bd41d1ee7f279).

## Browse the data

- `verified_sites`: 48 verified model/SAE layer entries, exact hooks, architecture, verification state and four recorded measurement fields.
- `metric_results`: 192 numeric rows (four measured metrics per site): mean L0, reconstruction MSE, explained variance, and the logit L2 change for a selected feature-ablation condition.
- `source_manifest.json`: source revision, record counts, CSV SHA256 and immutable GitHub evidence locations.

The two CSV files contain actual values extracted from the frozen original per-site JSON reports. They are **not newly run experiments**. The original reports preserve additional per-metric sample counts, experiment configurations, token ledgers and run identifiers.

## Protocol

The recorded population evaluation uses **64 independent held-out prompt groups per site**, kept separate from feature selection. Prompts are based on recurring controlled templates. These results test software compatibility, benchmark execution, and reliability reporting under this protocol, **not general SAE interpretability or relative scientific quality**.

Some metric units and scales differ by model, hook, baseline or checkpoint. **Do not interpret cross-model comparisons as a benchmark ranking.** A site's `compatibility`, `benchmark`, and `reliability` values represent completed acceptance of the recorded engineering protocol.

## Provenance and original evidence

- [Verified layer matrix](https://github.com/OpenSAE/SAEBase/blob/main/docs/verified-model-matrix.md)
- [Immutable compact evidence](https://github.com/OpenSAE/SAEBase/releases/tag/stage-d-evidence-20261016)
- [Original raw run archives](https://github.com/OpenSAE/SAEBase/releases/tag/stage-d-original-raw-20261016)
- [Additional independent exploratory runs](https://github.com/OpenSAE/SAEBase/releases/tag/stage-d-recovery-03-experiments-20261016) (separate from this evaluation matrix)
- [SAEBase v0.3.0](https://github.com/OpenSAE/SAEBase/releases/tag/v0.3.0)

Individual measurements in this dataset can be cross-checked by opening the corresponding `source_report_url`. The `source_manifest.json` gives hashes for the two derived CSVs.

## Usage

```python
from datasets import load_dataset

sites = load_dataset("OpenSAE/SAEBase-Evaluation", "verified_sites", split="train")
metrics = load_dataset("OpenSAE/SAEBase-Evaluation", "metric_results", split="train")
print(len(sites), len(metrics))  # 48, 192
```

See the [SAEBase Research Lab](https://huggingface.co/spaces/OpenSAE/SAEBase) for an interactive exploration of the same records.

## Distribution scope

This dataset contains derived records and links, **not** original model or SAE weight binaries, private activations, restricted source datasets, or third-party checkpoints. The original archives remain at the GitHub evidence releases. The project software uses Apache-2.0; no separate license for third-party datasets or checkpoint contents is asserted by this card. Review redistribution terms before adding any non-original assets.