A 31k-parameter transformer trained on synthetic data predicts real blood glucose zero-shot
0xdeadf1sh · reddit · 2026-10-05
A T1D developer extended his encoder-only transformer for blood sugar prediction by training it purely on outputs from his own T1DM patient simulator, then zero-shot testing on real CGM traces.
- Just 31,251 parameters (16 layers, 1 attention head each, hidden dim 16); trained in under 60 minutes on an NVIDIA DGX Spark
- Predicts the next 2 hours, usable autoregressively for 8-hour nocturnal forecasts, with counterfactual reasoning baked in
- Tested via his Android app (ExecuTorch backend) on 30 days of traces from Libre 3 Plus, Anytime CT5, and Linx sensors
- LoRA adapters in the app handle light personalization, but all published figures come from the base model
Model, simulator, and Android app are all open source.
More from Embodied
- FailBank Turns Runtime Shield Feedback into Persistent VLA Policy Gains (+25.4 Points Success) — notredame · 2026-10-05
- Reka releases Rho-1 research preview: a 19B omni model unifying text, images, video and robot actions — RekaAILabs · 2026-10-05
- NVIDIA's ENPIRE lets coding agents self-improve robot policies in the real world, hitting 99% success — chris_j_paxton · 2026-10-05
- Sensori demos humanoid Yuri: natural voice chat plus VR and exoskeleton teleoperation — YuXiang_IRVL · 2026-10-05
- UniWAM unifies world model and action policy, hits SOTA on LIBERO and RoboTwin benchmarks — LexiLove · 2026-10-05
- EyeRobot 2.0: active gaze lifts robot success from 52% to 67% — no wrist cameras needed — stepjamUK · 2026-10-05