CSC490 fall detection: trained model files
Code: https://github.com/soloadio/CSC490-Fall-Detection-Capstone-Project
Base models are not stored here; download them from Hugging Face (google/gemma-3-12b-it is gated: accept its terms first).
| Folder | What | Base model | Used by |
|---|---|---|---|
videomae/ |
VideoMAE-base fine-tuned for the falling motion (EDF + MCFD + UP-Fall). Full model, not an adapter | (from MCG-NJU/videomae-base-finetuned-kinetics) |
the Space (live detector) |
gemma/adapter_reasons/ |
LoRA r16, trained on verdict + evidence sentence (2026-11-08). Current checker | google/gemma-3-12b-it |
Modal checker (app/modal_app.py) |
gemma/adapter_v1/ |
LoRA r16, trained on the verdict word only. Previous checker, kept for rollback | google/gemma-3-12b-it |
- |
qwen/r16_all3/ |
LoRA r16 on the language layers, verdict only (Solomon; copy of Soloadio/qwen2vl-7b-fall-lora) |
Qwen/Qwen2-VL-7B-Instruct |
Modal checker (QwenVerifier) |
Load an adapter:
from peft import PeftModel
from transformers import Gemma3ForConditionalGeneration
model = Gemma3ForConditionalGeneration.from_pretrained("google/gemma-3-12b-it", dtype="bfloat16")
model = PeftModel.from_pretrained(model, "CSC490/fall-models", subfolder="gemma/adapter_reasons")
Results: research/results/ in the GitHub repo. A copy of these files is also on the Alliance Fir cluster in
/project/6010122/group12/models/.
Licences and intended use
Research prototype from a university course project. Not a medical device; do not rely on it to protect anyone. Each folder keeps the licence of what it was built from, and the training data is for non-commercial use, so use these files for non-commercial research and teaching only.
| Folder | Built from | Licence / terms |
|---|---|---|
videomae/ |
MCG-NJU/videomae-base-finetuned-kinetics | CC BY-NC 4.0 (non-commercial, credit the VideoMAE authors) |
gemma/* |
google/gemma-3-12b-it | Gemma Terms of Use and its Prohibited Use Policy apply to these adapters and to anything made with them |
qwen/* |
Qwen/Qwen2-VL-7B-Instruct | Apache 2.0 |
Training data: OmniFall annotations of the EDF, MCFD (Multiple Cameras Fall Dataset) and UP-Fall recordings; the annotations and the recordings belong to their authors and are released for non-commercial research (see the OmniFall dataset card and each dataset's terms).