Dev trains a 31K-parameter LoRA-tuned transformer to predict his blood sugar 2 hours ahead

0xdeadf1sh · reddit · 2026-11-14

Reddit user 0xdeadf1sh open-sourced T1DMAI: a tiny 31,253-parameter encoder-only transformer that predicts his own blood glucose.

Training: pretraining on 19,200 CGM-years of synthetic data (1 million simulated type-1 diabetics, 7 days each), then LoRA fine-tuning on 115 days of his own CGM data (80 train / 35 test).
Inputs/outputs: CGM time series plus insulin action curves (exact units) and carb curves (grams + glycemic index, estimated); trained in Kovatchev risk-space with {DILATE, MSE} + pinball loss and learned Kendall-Gal weights.
Capabilities: 2-hour-ahead prediction, missing-trace infilling, backcasts; time-of-day prediction emerges at >=48h context (native window 3.5–7 days).
Limits: still worse than the author at predicting hypoglycemia within 30 minutes; counterfactual reasoning lacks a downstream benchmark.

Source code is on GitHub.

More from Research