SAIL — Pretrained SAE Weights

Pretrained Matryoshka Sparse Autoencoder (SAE) weights for the SAIL repository. See the project page for the full pipeline and usage instructions.

Two checkpoints are provided, one for each foundation model (FM) embedding space:

File	Foundation model	Input dim	Dictionary sizes	k values
biomedparse_sae.ckpt	BiomedParse	1536	128, 512, 2048, 8192	20, 40, 80, 160
dinov3_sae.ckpt	DINOv3	1024	128, 512, 2048, 8192	5, 10, 20, 40

Both SAEs were trained on CT and MRI embeddings from the TotalSegmentator dataset.

Usage

To download these weights and place them in the expected directory structure, run from the repo root:

bash pretrained/download_weights.sh

Citation

If you find this work useful, please cite our paper:

  @misc{sail2026,
    title = {Sparse Autoencoders for Interpretable Medical Image Representation Learning},
    author = {Wesp, Philipp and Holland, Robbie and Sideri-Lampretsa, Vasiliki and Gatidis, Sergios},
    year = 2026,
    journal = {arXiv.org},
    howpublished = {https://arxiv.org/abs/2603.23794v1}
  }