Brunosil22/Orena_FRAME_Track_Qwen3_VL_8B_finetuned

ICVS-2Ai ORena FOCUS — FRAME

Qwen3-VL-8B adapter for single-frame surgical foreign-object VQA. The selected public checkpoint is FRAME20K step 3300, trained at 768 px with vision/language LoRA and image augmentation. No additional annotations were used. The available records show that the last verified platform result used the earlier CV16K step-3258 checkpoint, so step 3300 is not labelled here as a verified submission.

Files and flow

• —adapter_step_03300/: final PEFT configuration and weights.
• —inference.py: challenge entry point; loads one PNG, resizes it to 768 px, applies the frozen timestamped prompt, generates greedily, and normalizes strict answer formats.
• —resources/system_prompt.txt: exact training and inference prompt.

For the intended merged deployment, merge the adapter into Qwen/Qwen3-VL-8B-Instruct. For adapter-based execution, set MODEL_DIR to the base model and ADAPTER_DIR to adapter_step_03300, then provide FOCUS request.json and frames/<qID>.png under INPUT_PATH:

MODEL_DIR=Qwen/Qwen3-VL-8B-Instruct \
ADAPTER_DIR=adapter_step_03300 INPUT_PATH=/input OUTPUT_PATH=/output \
python inference.py

Training code

The shared, uniform-sampling training release is available at Brunosil22/orena-focus-training. The FRAME recipe is `configs/frame_step3300.env`; FRAME uses the single annotated frame and has no learned frame-selector dependency.