🛡️ Fall / Danger Detection — single VLM

A Qwen3.5-2B fine-tune that flags falls, a person down, or distress from a short clip — trained mostly on synthetic 3D data. Chosen for capability on hard real cases over raw benchmark recall — see the model card for the honest trade-off vs. the SmolVLM2 sibling model. Code · Model card.

Try an example (real footage, UR Fall Detection dataset [Kwolek & Kepski, 2014]) or upload a clip. Trained on synthetic data, so it may false-alarm on out-of-distribution real scenes. Research prototype — not a medical/safety device.