Downloads · 30 days
4
7% of all-time downloads
freuid-2026-mangojump
ching0206/freuid-2026-mangojump
freuid-2026-mangojump is a image classification model from ching0206. Use it when you need a label for an image. The card lists the license as mit.
Frozen weights for our FREUID Challenge 2026 (IJCAI-ECAI) submission: feature-level fusion of DoRA-adapted foundation backbones with an OOD-routed presentation-attack (PAD) head. Code, Docker contract, and the technic…
All-time downloads
61
Public
Repo size
2.1 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.npz1.5 GB · 74%
At a glance
- Task
- Image Classification
- License
- mit
- Access
- Public
- Created
- Jul 10, 2026
- Updated
- Jul 12, 2026
- SHA
- 8c145f9e
Model card
From the Hugging Face model README
FREUID 2026 — team mangojump frozen inference system
Frozen weights for our FREUID Challenge 2026 (IJCAI-ECAI) submission:
feature-level fusion of DoRA-adapted foundation backbones with an
OOD-routed presentation-attack (PAD) head. Code, Docker contract, and the
technical report live in the public repository (see the competition reply);
this repo holds only the weight artifacts consumed by src/predict_docker.py.
Contents
| File | What it is |
|---|---|
<member>.pt (dino, dino_hplus, siglip512, dfn5b, dino_hplus_dlc, dino_hplus_ds) | DoRA adapter deltas (rank 16, α 32) + EMA copies per member; base backbone weights are NOT included |
heads.pt | fusion head, capture head, PAD head (linear, LayerNorm→Dropout→Linear) |
fisher_idx.npz | frozen FGTS top-64 token indices for DINOv3 members |
knn_ref.npz | block-normalized digital-train reference matrix (fp32) for the kNN router |
config.json | member specs, routing thresholds (capture 0.5, distance floor 0.246778), variant map |
Base models (fetch from original sources; only deltas are hosted here)
- DINOv3 ViT-L/16 & ViT-H+/16 (
timm/vit_{large,huge_plus}_patch16_dinov3.lvd1689m) — Meta AI, DINOv3 license (timm mirrors, ungated) - SigLIP-2 SO400M/16 @512 (
timm/vit_so400m_patch16_siglip_512.v2_webli) — Apache-2.0 - DFN5B CLIP ViT-H/14 @378 (
timm/vit_huge_patch14_clip_378.dfn5b) — Apple ASCL
Training data of the adapters
FREUID 2026 competition data; two members additionally saw small fractions of DLC-2021 (CC BY-SA 2.5) and SIDTD (CC BY-SA 4.0). Neither dataset is redistributed here. Adapter deltas are released under MIT; base-model and dataset licenses continue to apply to their respective artifacts.
- Task
- Image Classification
- License
- mit
- Created
- Jul 10, 2026
- Updated
- Jul 12, 2026