HC-MTS production models
Checkpoints for the HC-MTS consultancy stack (GitHub).
Files
| Path | Description |
|---|---|
l1/s2_convnext/best.pt |
3-way clinical triage (ConvNeXt-Base) |
l15/s2_l15_natural/best.pt |
10-family rash head (conditional) |
seg/s0_att_resunet/best.pt |
Lesion segmentation (AttResUNet) |
face_guard/*.onnx |
Face parser + YuNet detector for mask cleanup |
export/inference_config.json |
Temperature, abstain, class names |
research/* |
Additional training checkpoints (S1, S3, L1.5 balanced, wound copy) |
Load in Python
from pathlib import Path
import torch
from hc_mts.models.convnext_dx import ConvNeXtDx
root = Path("runs/s2_convnext/best.pt") # after scripts/setup_from_hf.sh
ckpt = torch.load(root, map_location="cpu", weights_only=False)
model = ConvNeXtDx(num_classes=len(ckpt["classes"]), pretrained=False)
model.load_state_dict(ckpt["model"])
Metrics (test)
- L1: accuracy 0.854, macro-F1 0.836
- L1.5: macro-F1 0.479 (family level)
- Seg: Dice 0.898
See dataset repo metrics/ for full JSON reports.