MICCAI ORena 2026 โ SAVE FOCUS submissions
Qwen3.5-4B + LoRA (r=64, ฮฑ=128) models for the ORena SAVE FOCUS challenge.
| Folder | Track | Adapter |
|---|---|---|
frame/ |
Frame | capability-rebalanced anneal, step 2428 |
segment_procedure/ |
Segment and Procedure (one model, routed by window length: โค 305 s = segment) | unified video, epoch 8 |
base_model/ |
shared | unmodified Qwen/Qwen3.5-4B |
Run
Requires a CUDA GPU and Python 3.11. Each inference.py is the script from the submitted container: it reads /input and writes /output/answer.json.
hf download QianyeYang/MICCAIORena26 --local-dir orena26 && cd orena26
ln -s ../../base_model frame/resources/model
ln -s ../../base_model segment_procedure/resources/base_model
pip install -r frame/requirements.txt # or segment_procedure/requirements.txt (torch wheel note inside)
export FOCUS_VENDOR_PATH=$(mktemp -d); touch $FOCUS_VENDOR_PATH/.complete # use installed packages instead of the container's vendor archive
python frame/inference.py # or segment_procedure/inference.py
Input: /input/request.json, /input/FO_definitions.json, and /input/frames/<qID>.png (frame) or /input/plain/<qID>.mp4 (segment/procedure).
- Downloads last month
- -