OpenSLA
Sensor-Language-Action model (UCLA, Yuzhe Yang lab)
ECGPPGRespABPCO2EEGCGMvitalslabsText
Key facts
- Architecture
- Qwen3.5-2B + LoRA; frozen OSF-derived DINO 1D-ViT waveform encoder; hierarchical sensor memory (OpenSLA-H) or flat prefix (OpenSLA-B)
- Pretraining Data
- Sensor-language-action training on MC-MED, MIMIC-III, MOVER, VitalDB, MetaboNet, PEDAP (115,740 individuals across 7 datasets incl. MIMIC-IV external); waveform encoder from OSF DINO pretraining
- Venue
- arXiv 2026-10-06
Summary
It's not drop-in for Clarity. The domains are clinical/OR monitors (finger pleth, not wrist PPG) and CGM, with no ACC or wrist data. Re-check HF for the checkpoint.