SAEV
Collection
SAEs for vision models like CLIP or DINOv2 • 7 items • Updated • 5
SAE for Meta's DINOv3 ViT-L/16 trained on ImageNet-1K Activations
- Homepage: https://osu-nlp-group.github.io/saev
- Code: https://github.com/OSU-NLP-Group/saev
- Preprint: https://arxiv.org/abs/2511.17735
- Demos: https://osu-nlp-group.github.io/saev#demos
- Point of Contact: Sam Stevens
Checkpoints
Each checkpoint is a sparse autoencoder (SAE) trained on a different layer with a different sparsity level. Pick the checkpoint that matches your target layer and desired sparsity (L0).
| Run ID | Layer | L0 | MSE | Path |
|---|---|---|---|---|
| jsqj2arm | 13 | 2.8 | 4.9303 | layer_13/jsqj2arm/sae.pt |
| kk60aru4 | 13 | 6.3 | 4.6254 | layer_13/kk60aru4/sae.pt |
| y8q60ohz | 13 | 23.3 | 4.3962 | layer_13/y8q60ohz/sae.pt |
| ag8agm56 | 13 | 90.7 | 3.7613 | layer_13/ag8agm56/sae.pt |
| 220r8j1q | 13 | 247.5 | 3.6310 | layer_13/220r8j1q/sae.pt |
| fkl5sxba | 13 | 900.1 | 3.4952 | layer_13/fkl5sxba/sae.pt |
| fi7qafny | 15 | 1.2 | 12.0470 | layer_15/fi7qafny/sae.pt |
| txmrh5nd | 15 | 5.0 | 10.8379 | layer_15/txmrh5nd/sae.pt |
| aq8vvjub | 15 | 8.8 | 10.3427 | layer_15/aq8vvjub/sae.pt |
| xfgouwrz | 15 | 38.9 | 9.7837 | layer_15/xfgouwrz/sae.pt |
| rsmrpkly | 15 | 92.4 | 9.0052 | layer_15/rsmrpkly/sae.pt |
| e9oeml82 | 15 | 151.3 | 8.8066 | layer_15/e9oeml82/sae.pt |
| di427rrs | 17 | 1.3 | 27.7408 | layer_17/di427rrs/sae.pt |
| edx9q34f | 17 | 5.8 | 24.9920 | layer_17/edx9q34f/sae.pt |
| n7pv6rkj | 17 | 11.8 | 23.5505 | layer_17/n7pv6rkj/sae.pt |
| 4rhpmk3f | 17 | 42.6 | 22.8549 | layer_17/4rhpmk3f/sae.pt |
| jqx6qdxv | 17 | 141.7 | 20.8065 | layer_17/jqx6qdxv/sae.pt |
| vkdu21ck | 17 | 507.3 | 20.0385 | layer_17/vkdu21ck/sae.pt |
| y6osup5x | 19 | 3.0 | 75.8186 | layer_19/y6osup5x/sae.pt |
| aa30r3nm | 19 | 8.1 | 70.1213 | layer_19/aa30r3nm/sae.pt |
| 0tj48gqd | 19 | 22.2 | 64.6201 | layer_19/0tj48gqd/sae.pt |
| 7dr58kwn | 19 | 284.6 | 59.5570 | layer_19/7dr58kwn/sae.pt |
| 2uqtzyv6 | 19 | 418.0 | 57.5434 | layer_19/2uqtzyv6/sae.pt |
| s96104bm | 19 | 975.5 | 55.9320 | layer_19/s96104bm/sae.pt |
| qcyausyf | 21 | 2.2 | 215.4446 | layer_21/qcyausyf/sae.pt |
| x7py290w | 21 | 7.6 | 207.6037 | layer_21/x7py290w/sae.pt |
| v4pyroov | 21 | 28.1 | 186.8892 | layer_21/v4pyroov/sae.pt |
| t1ip1brk | 21 | 41.6 | 182.1836 | layer_21/t1ip1brk/sae.pt |
| pz4up9fd | 21 | 401.4 | 175.3216 | layer_21/pz4up9fd/sae.pt |
| y8vhxwya | 21 | 773.8 | 165.5911 | layer_21/y8vhxwya/sae.pt |
| lnleoyf6 | 23 | 0.0 | 1833.7889 | layer_23/lnleoyf6/sae.pt |
| t1vh0qy1 | 23 | 3.3 | 1657.9833 | layer_23/t1vh0qy1/sae.pt |
| fxcpfysr | 23 | 9.4 | 1503.3433 | layer_23/fxcpfysr/sae.pt |
| kd2pd8rs | 23 | 83.0 | 1446.8923 | layer_23/kd2pd8rs/sae.pt |
| 1qynjykb | 23 | 214.8 | 1347.6047 | layer_23/1qynjykb/sae.pt |
| 9fn4l6rf | 23 | 1574.5 | 1256.6656 | layer_23/9fn4l6rf/sae.pt |
This metadata is also available in manifest.jsonl at the repo root for programmatic access.
Usage
from huggingface_hub import hf_hub_download
import saev.nn
path = hf_hub_download("osunlp/SAE_DINOv3_ViT-L-16_IN1K", "layer_23/9fn4l6rf/sae.pt")
sae = saev.nn.load(path)
Inference Instructions
Follow the instructions here.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support