GestaltMatcher-Arc (GMDB) β self-trained reproduction
ArcFace ResNet-50 face encoder, pre-trained on Glint360K (~360K identities, ~17M faces) and fine-tuned on GestaltMatcherDB (GMDB) v1.1.2 to embed the facial gestalt of rare genetic syndromes. Produces a 512-d embedding; diagnosis is nearest-prototype cosine retrieval.
This is the face-specialist component of the SyndroDX open-world rare-disease diagnosis framework.
Training
| backbone | ArcFace ResNet-50 (Glint360K pre-trained) |
| fine-tune data | GMDB v1.1.2, original (non-enhanced) aligned faces |
| loss | ArcFace margin |
| image size | 112 x 112 |
| embedding dim | 512 |
| batch size | 2048 |
| epochs | 50 |
Benchmark (GMDB official patient-level split, n=476)
| metric | value |
|---|---|
| top-1 | 52.5% |
| top-5 | 66.8% |
| top-10 | 74.2% |
Provenance, license & responsible use
- Derived from GestaltMatcher-Arc (CC-BY-NC 4.0) β non-commercial, attribution required.
- Fine-tuned on GMDB, an access-controlled database of patient facial images. These weights are derived from controlled patient data; do not redistribute publicly without confirming the GMDB data use agreement. Private / research use.
- Face-recognition model trained on patient faces β handle under appropriate privacy/ethics controls.
Cite GestaltMatcher / GestaltMatcher-Arc and GMDB when using this model.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support