edge-fall-vlm-256m — 256M fall / danger detector (size-comparison sibling)
A SmolVLM2-256M-Video-Instruct fine-tune for fall / person-down / distress detection, same recipe as edge-fall-vlm-2.2b but at 256M. Published mainly to show how accuracy scales with VLM size.
Code: https://github.com/vieenrose/edge-fall-vlm · Demo (pick a size): https://huggingface.co/spaces/Luigi/edge-fall-vlm-demo
Accuracy vs size (same recipe, same real test sets)
| Model | URFD (easy) recall/spec | OOPS in-the-wild recall |
|---|---|---|
| 256M | 1.0 / 1.0 | 0.13 |
| 500M | 1.0 / 1.0 | 0.31 |
| 2.2B (recommended) | 0.90 / 1.0 | 0.83 |
This 256M model saturates the easy in-distribution test but MISSES most real in-the-wild falls (recall 0.13). Smaller VLMs fit the training distribution but do not generalize to novel real footage. For the actual safety task, use the 2.2B. This model is useful for research / the fastest possible on-device path where recall is not safety-critical.
Files: transformers model.safetensors + model-Q8_0.gguf / mmproj-f16.gguf
(llama.cpp). Apache-2.0. Not a medical/safety device.