vthawfeek/mtdna-fm-haplogroup

mtDNA-FM Haplogroup Adapter (LoRA r=8)

LoRA adapter for haplogroup classification (26 major haplogroups) on top of vthawfeek/mtdna-foundation-model.

Usage

from mtdna_fm.model.model import MtDNAForHaplogroupClassification, MtDNAModel
from peft import PeftModel

base = MtDNAModel.from_pretrained("vthawfeek/mtdna-foundation-model")
model = MtDNAForHaplogroupClassification(base, num_labels=26)
model = PeftModel.from_pretrained(model, "vthawfeek/mtdna-fm-haplogroup")
model.eval()

LoRA Configuration

• r = 8, lora_alpha = 16
• target_modules: query, key, value, dense
• lora_dropout = 0.1

Task

26-class haplogroup classification: A, B, C, D, E, F, G, H, HV, I, J, K, L0, L1, L2, L3, L4, L5, M, N, R, T, U, V, W, X.

Input: full 16,569-bp mtDNA genome sequence. Mean-pooled CLS embeddings across overlapping windows (size=512, stride=256) fed into a Linear(256, 26) head.

Limitations

HmtDB training data has a European population bias (haplogroup H is overrepresented). Performance on underrepresented African L sub-haplogroups may be lower.