Models for more layers

[krishnakanthnakkav2]

Hello Authors,

Thank you for releasing the codebase. Do you plan to release the SAEs at different layers of the backbone model?

Regards

[samuelstevens]

We do not. However, we plan on training additional SAEs with more modern activations (JumpReLU, BatchedTopK) on various ViTs (SigLIP 2, larger variants of DINOv2, etc). Are there particular model/dataset/layer combinations that you are interested in? If so, let me know and I can be sure to include them in my upcoming jobs.

[marcantoinewilly]

Would it be possible for you to provide a pre-trained SAE checkpoint for CLIP on CelebA in the near future?