README.md

---
library_name: saelens
---

SAE panels and SAEBench results from the paper "[Are Sparse Autoencoder Benchmarks Reliable?](https://arxiv.org/abs/2605.18229)"

This repo is split into 2 panels, a cross-architecture panel consisting of 4 SAEs (K=50 Matryoska, k=100 Matryoshka, k=50 BatchTopK, k=100 BatchTopK),
and a Matryoshka panel consisting of 4 Matryoshka SAEs verying the number of Matryoshka prefixes from 1 to 4 (n-1, n-2, n-3, n-4). Each SAE in the Matryoshka panel
is trained 3 times with different seeds (so 12 SAEs total). The cross-architecture panel is trained for 1.5B tokens, while the Matryoshka panel is trained for 300M tokens.

Within each SAE dir, there are a number of snapshots of the SAE taken throughout training. Each of these snapshot dirs include the following:

- SAE weights (`sae_weights.safetensors`) and `cfg.json` for loading with SAELens
- SAEBench raw result JSON files for all SAEBench metrics

To load an SAE snapshot using SAELens, run the following:

```python
from sae_lens import SAE

sae = SAE.from_pretrained("decoderesearch/sae-snapshot-panels", "path/to/snapshot")
```

For instance, to load the SAE snapshot for the K=100 BatchTopK SAE after 500M tokens of training, you would run:

```python
sae = SAE.from_pretrained(
  "decoderesearch/sae-snapshot-panels",
  "cross-arch-panel/gemma-2-2b/batchtopk/k-100/seed-0/snapshots/step-122070-tokens-500000000",
)
```

## Citation

If you use these SAEs in your work, please cite the following:

```bibtex
@misc{chanin2026saebenchmarks,
      title={Are Sparse Autoencoder Benchmarks Reliable?}, 
      author={David Chanin},
      year={2026},
      eprint={2605.18229},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2605.18229}, 
}
```