GlucoseDao/glumind-global-h12

GluMind is a Transformer-based glucose forecasting model that predicts future blood glucose values from CGM readings combined with heart rate and step-count features.

Model Details

Parameter	Value
Horizon	12 steps (×5 min each)
Input steps	80
d_model	32
n_heads	4
n_blocks	3
ff_units	128
dropout	0.1
Training mode	global
Split scheme	classic
Best epoch	26
Best val loss	0.0015516681092495107

Validation Metrics (overall)

mae,rmse,mard
11.433954238891602,17.869876861572266,8.46254825592041

Test Metrics (overall)

mae,rmse,mard
11.698744773864746,18.456811904907227,8.520216941833496

Usage

Download the model and evaluate with:

uv run scripts/glumind/download_from_huggingface.py \
    --repo-id GlucoseDao/glumind-global-h12 \
    --output-dir test_model

uv run scripts/glumind/evaluate_glumind.py \
    --run-dir test_model \
    --test-csv test_data/livia_glumind_ready.csv

Files

File	Description
best_model.pt	Best model weights (PyTorch state dict)
config.json	Full training configuration
tuning_meta.json	Training metadata including dataset stats
best_info.json	Best epoch and validation loss
tuning.txt	Training log
*_metrics_*.csv	Validation and test metrics breakdowns