You can deploy this model and start earning money today!

osunlp/
SAE_DINOv3_ViT-S-16_IN1K

2511.17735

mit

Views

No views yet

SAE for Meta's DINOv3 ViT-S/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
3zih0tpa	6	0.6	1.4566	`layer_6/3zih0tpa/sae.pt`
f5qog0he	6	4.9	1.1679	`layer_6/f5qog0he/sae.pt`
u5aqv7t7	6	19.4	1.0534	`layer_6/u5aqv7t7/sae.pt`
usiwpodi	6	45.3	0.9598	`layer_6/usiwpodi/sae.pt`
nbqbvh45	6	193.2	0.8631	`layer_6/nbqbvh45/sae.pt`
sfpco1tn	6	431.9	0.8223	`layer_6/sfpco1tn/sae.pt`
l9stkmwt	7	0.7	2.8639	`layer_7/l9stkmwt/sae.pt`
tbfdr3cc	7	4.6	2.3978	`layer_7/tbfdr3cc/sae.pt`
zaxl9nqu	7	9.3	2.2369	`layer_7/zaxl9nqu/sae.pt`
r2g7cj5v	7	26.9	2.1348	`layer_7/r2g7cj5v/sae.pt`
d5ej9yuh	7	102.6	1.8223	`layer_7/d5ej9yuh/sae.pt`
36538viq	7	322.3	1.7001	`layer_7/36538viq/sae.pt`
p0z7t1ci	8	1.9	5.3336	`layer_8/p0z7t1ci/sae.pt`
125xh1t9	8	5.2	4.9238	`layer_8/125xh1t9/sae.pt`
qwf07reo	8	12.3	4.4735	`layer_8/qwf07reo/sae.pt`
tzwgex0i	8	48.4	4.2854	`layer_8/tzwgex0i/sae.pt`
5gjy7lwi	8	159.8	3.7245	`layer_8/5gjy7lwi/sae.pt`
bvsb2257	8	352.9	3.5116	`layer_8/bvsb2257/sae.pt`
qt0fmmxm	9	7.4	11.2642	`layer_9/qt0fmmxm/sae.pt`
1o4uc5bf	9	17.5	10.1970	`layer_9/1o4uc5bf/sae.pt`
z1qvy51u	9	84.1	10.1741	`layer_9/z1qvy51u/sae.pt`
1ihxsv0i	9	167.0	8.9344	`layer_9/1ihxsv0i/sae.pt`
euoj6wv0	9	229.1	8.7690	`layer_9/euoj6wv0/sae.pt`
ickedctl	9	381.9	8.4021	`layer_9/ickedctl/sae.pt`
chn5wi3x	10	1.8	28.7944	`layer_10/chn5wi3x/sae.pt`
219r3phu	10	11.0	25.1447	`layer_10/219r3phu/sae.pt`
knglrhzb	10	23.6	23.1152	`layer_10/knglrhzb/sae.pt`
jt45lucm	10	29.7	22.6514	`layer_10/jt45lucm/sae.pt`
6hrok1al	10	234.3	20.5268	`layer_10/6hrok1al/sae.pt`
g4dexqq1	10	387.2	19.6115	`layer_10/g4dexqq1/sae.pt`
jgu19fzx	11	4.3	279.2303	`layer_11/jgu19fzx/sae.pt`
8yd05vxi	11	6.9	263.3258	`layer_11/8yd05vxi/sae.pt`
utmjp20e	11	36.2	253.5636	`layer_11/utmjp20e/sae.pt`
gc6iqrf2	11	95.1	240.0467	`layer_11/gc6iqrf2/sae.pt`
5ewxrjg4	11	130.8	222.1000	`layer_11/5ewxrjg4/sae.pt`
36ztscy4	11	420.3	206.9659	`layer_11/36ztscy4/sae.pt`

This metadata is also available in manifest.jsonl at the repo root for programmatic access.

Usage

python

from huggingface_hub import hf_hub_download

import saev.nn

path = hf_hub_download("osunlp/SAE_DINOv3_ViT-S-16_IN1K", "layer_11/36ztscy4/sae.pt")
sae = saev.nn.load(path)

Inference Instructions

Follow the instructions here.