HyenaLAI — Local Ancestry Inference weights
Pretrained checkpoints for HyenaLAI, a single subquadratic Hyena model that performs Local Ancestry Inference from SNPs. See the code repository for architecture, training and the zero-setup demo notebook.
Contents
One subfolder per (<classes>/<window>) configuration; each ships classifier.ckpt
(Stage-2 fine-tuned model) and config.yaml (model kwargs to rebuild it without Hydra).
examples/ holds a few per-individual HDF5s (v3) so the demo runs with no preprocessing.
| Task | Window (SNPs) | Val accuracy |
|---|---|---|
| 3 classes | 1024 | 97.03% |
| 3 classes | 4096 | 99.50% |
| 3 classes | 32768 | 99.89% |
| 5 classes | 1024 | 77.93% |
| 5 classes | 2048 | 96.12% |
| 5 classes | 4096 | 95.28% |
Ancestry classes and their populations
Each class groups reference populations from the 1000 Genomes Project (and, for NAT_PURE, the
Human Genome Diversity Project):
| Class | Description | Populations (source panel) |
|---|---|---|
| EUR | European | GBR (British), CEU (Utah, CEPH European) |
| AFR | African | YRI (Yoruba), ESN (Esan), MSL (Mende), GWD (Gambian Mandinka) |
| EAS | East Asian | CHS (Southern Han), JPT (Japanese), CDX (Dai Chinese), CHB (Han Chinese), KHV (Kinh Vietnamese) |
| SAS | South Asian | ITU (Telugu), STU (Sri Lankan Tamil), GIH (Gujarati), PJL (Punjabi), BEB (Bengali) |
| NAT | Native American (admixed 1000G) | PEL (Peruvian), MXL (Mexican ancestry) |
| NAT_PURE | Native American (unadmixed HGDP) | Maya, Pima, Karitiana, Surui, Colombian |
3-class models use EUR / AFR / EAS. Among the 5-class models, the 1024 and 4096 windows
use EUR / AFR / EAS / NAT / SAS, while the 2048 window uses the same set but with
NAT_PURE (unadmixed HGDP Native American references) in place of NAT. Always check a model's
config.yaml locations field for the exact class order.
Usage
from huggingface_hub import hf_hub_download
ckpt = hf_hub_download("gcameto/hyena-lai", "3class/4096/classifier.ckpt")
config = hf_hub_download("gcameto/hyena-lai", "3class/4096/config.yaml")
See notebooks/demo_inference.ipynb in the code repo for end-to-end whole-genome inference.
License & citation
Apache 2.0. If you use these weights, please cite the HyenaLAI paper (CIARP 2026) and thesis (UdelaR, 2026) — see the code repository for BibTeX.