DINOv3-SURGE (Base)
A DINOv3 ViT-B/16 vision encoder adapted to the surgical / endoscopic domain, used as an additively-fused second vision encoder in a Qwen3.5-9B video-QA model for the ORena SAVE FOCUS challenge (MICCAI 2026), PROCEDURE track.
Access is gated: the weights are DINOv3-derived, and the DINOv3 License Agreement travels with them.
Files
| File | Size | SHA256 |
|---|---|---|
SURGE_base.pth | 527,391,215 B | fdeac7df8e2eacb935853198ff397f6e4254bd11523718ba9d20f66c21afe0ac |
Always check the hash after downloading — the consuming code does, and refuses to train against a checkpoint that does not match.
Architecture
Verified against the checkpoint itself, not assumed:
| Architecture | vit_base (DINOv3 DinoVisionTransformer) |
| Parameters | 85,669,632 |
| Depth / embed dim / heads | 12 / 768 / 12 |
| Patch size | 16 |
| FFN | mlp, ratio 4.0 |
| Storage (register) tokens | 4 |
| Norm layer | layernorm, layerscale init 1e-5 |
| RoPE | base 100.0, normalize_coords="separate" |
| State-dict branch | teacher |
At 448x448 input, forward_features returns patch tokens of shape (1, 784, 768).
The checkpoint is a plain training checkpoint, so the backbone weights live under a
teacher key and carry module. / backbone. / feature_model. prefixes. A loader
must select the teacher branch (falling back to model, then the object itself) and
strip those prefixes.
Usage
The architecture code is not in transformers; it is upstream DINOv3
(facebookresearch/dinov3, commit 6876159a11b4df116f30f667f8c9888617df0751).
PROVENANCE.txt in this repository records exactly which files were used and the two
respects in which the supplied vision_transformer.py differs from upstream.
from huggingface_hub import hf_hub_download
path = hf_hub_download("ethanshili/dinov3-SURGE", "SURGE_base.pth")
Instantiate DinoVisionTransformer directly with the config above rather than through
the vit_base() factory, which hardcodes its own embed_dim/depth/num_heads/
ffn_ratio.
A complete, working loader — checkpoint key resolution, LoRA adaptation, and fusion into
a Qwen vision stream — is in the focus_v15.fusion module of the accompanying pipeline
repository.
Provenance and what is not documented here
PROVENANCE.txt covers the vendored DINOv3 architecture code. The SURGE adaptation
itself — the surgical corpus it was trained on and its training recipe — is not
documented in this repository; these weights were received as a finished checkpoint.
One consequence matters in practice: the normalization SURGE was trained under is not
recoverable from the checkpoint, so downstream code that feeds it ImageNet-normalized
frames is relying on an assumption, not a verified fact.
License
DINOv3 License Agreement — see DINOV3_LICENSE.md. Section 1(b) permits redistribution
only under the terms of that agreement and requires a copy of it to accompany any copy
of the materials. Keep DINOV3_LICENSE.md and PROVENANCE.txt with these weights.