SAEV
Collection
SAEs for vision models like CLIP or DINOv2 • 7 items • Updated • 5
SAE for Meta's DINOv3 ViT-B/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 |
|---|---|---|---|---|
| y7uk853s | 6 | 2.2 | 2.8623 | layer_6/y7uk853s/sae.pt |
| e4i667sz | 6 | 5.1 | 2.6635 | layer_6/e4i667sz/sae.pt |
| extgn8yv | 6 | 19.5 | 2.5258 | layer_6/extgn8yv/sae.pt |
| 8hyzbyht | 6 | 79.1 | 2.1732 | layer_6/8hyzbyht/sae.pt |
| t2xcozei | 6 | 277.5 | 2.1249 | layer_6/t2xcozei/sae.pt |
| lppv40ws | 6 | 725.8 | 1.9904 | layer_6/lppv40ws/sae.pt |
| iyb7ec1w | 7 | 0.1 | 7.7452 | layer_7/iyb7ec1w/sae.pt |
| 688bm8ht | 7 | 4.4 | 6.3078 | layer_7/688bm8ht/sae.pt |
| 40j71aj2 | 7 | 8.7 | 6.0463 | layer_7/40j71aj2/sae.pt |
| xucm378k | 7 | 35.4 | 5.6832 | layer_7/xucm378k/sae.pt |
| knz7yndg | 7 | 132.6 | 5.0626 | layer_7/knz7yndg/sae.pt |
| 6yhupj05 | 7 | 598.6 | 4.7180 | layer_7/6yhupj05/sae.pt |
| wgh9hgih | 8 | 0.5 | 19.4129 | layer_8/wgh9hgih/sae.pt |
| ttghd72n | 8 | 7.1 | 16.0403 | layer_8/ttghd72n/sae.pt |
| q6pg7hl9 | 8 | 12.2 | 15.1630 | layer_8/q6pg7hl9/sae.pt |
| r3opp7dy | 8 | 63.1 | 14.6620 | layer_8/r3opp7dy/sae.pt |
| bk7iwhfu | 8 | 215.1 | 13.1875 | layer_8/bk7iwhfu/sae.pt |
| rc82kpln | 8 | 690.9 | 12.5601 | layer_8/rc82kpln/sae.pt |
| cozptrw2 | 9 | 1.1 | 45.2121 | layer_9/cozptrw2/sae.pt |
| 1aod3v62 | 9 | 10.7 | 39.1680 | layer_9/1aod3v62/sae.pt |
| zvx4qkov | 9 | 20.1 | 36.4623 | layer_9/zvx4qkov/sae.pt |
| na2k2dyp | 9 | 84.4 | 36.4057 | layer_9/na2k2dyp/sae.pt |
| tqk8igwb | 9 | 332.4 | 32.2874 | layer_9/tqk8igwb/sae.pt |
| 893a4vol | 9 | 963.5 | 31.2656 | layer_9/893a4vol/sae.pt |
| eqht2edc | 10 | 8.7 | 132.2482 | layer_10/eqht2edc/sae.pt |
| 1hjlnu1s | 10 | 21.5 | 115.6000 | layer_10/1hjlnu1s/sae.pt |
| jpnwfh3w | 10 | 47.5 | 108.0820 | layer_10/jpnwfh3w/sae.pt |
| 2we45xxf | 10 | 341.3 | 99.9972 | layer_10/2we45xxf/sae.pt |
| bv1h09se | 10 | 454.0 | 98.0921 | layer_10/bv1h09se/sae.pt |
| jjewtqwp | 10 | 971.8 | 97.2687 | layer_10/jjewtqwp/sae.pt |
| n1xwev0z | 11 | 5.0 | 640.6468 | layer_11/n1xwev0z/sae.pt |
| ef657fwa | 11 | 6.5 | 621.5192 | layer_11/ef657fwa/sae.pt |
| qoc1660r | 11 | 89.9 | 560.8286 | layer_11/qoc1660r/sae.pt |
| 6crsj9gj | 11 | 123.8 | 547.8894 | layer_11/6crsj9gj/sae.pt |
| d4v8aruu | 11 | 435.5 | 532.5820 | layer_11/d4v8aruu/sae.pt |
| 22p3bnt8 | 11 | 786.3 | 508.5309 | layer_11/22p3bnt8/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-B-16_IN1K", "layer_11/22p3bnt8/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