opscribe-ai/surgvu26-cat2-v6.2

SurgVU 2026 Category 2 — submitted weights

Every learned artifact in our SurgVU 2026 Category 2 submission.

Files

file	size	what it is
tools_resnet50_long.pt	91 MB	ResNet-50, 12 instrument classes, multi-label. Predicts install state over a 30 s window. macro-F1 0.7802
task_resnet50_long.pt	91 MB	ResNet-50, 8 activity classes, softmax. accuracy 0.9348
vlm_lora/	364 MB	LoRA r=32, alpha=64, for Qwen2.5-VL-7B. 95.2M of 4.79B parameters
variant_head.pt	43 MB	instrument size family, large vs mega
yolo_best.pt	14 MB	YOLOv5-small instrument detector, fine-tuned 100 epochs at 640px on 14 SurgVU instrument classes (two of which the challenge rules out of scope)

Training data

• opscribe-ai/surgvu-cat2-vqa — 23,354 QA pairs, used for the LoRA's second stage
• opscribe-ai/mis-abdominal-gi-min-invasive — 86,991 pairs from nine public corpora, the LoRA's first stage
• The perception models, detector and variant head were trained only on the published SurgVU dataset (CC BY 4.0)

The 11 publicly graded sample cases (case_122–case_132) were excluded

The LoRA's base

Qwen/Qwen2.5-VL-7B-Instruct → nvidia/Qwen2.5-VL-7B-Surg-CholecT50 → our LoRA. Load the NVIDIA checkpoint and apply vlm_lora/ on top. Served NF4-quantised, 16 frames.

Licences — per file, not uniform

file	licence
the two ResNet-50s, variant_head.pt	torchvision ImageNet initialisation (BSD-3-Clause) fine-tuned on CC BY 4.0 data
vlm_lora/	inherits from nvidia/Qwen2.5-VL-7B-Surg-CholecT50, whose card lists its licence as "other" — check it before redistributing
yolo_best.pt	Fine-tuned from yolov5s.pt (COCO-pretrained YOLOv5-small) using the YOLOv5 codebase, which is GPL-3.0. A copy of that codebase ships inside the submission container, so the container carries GPL-3.0 code; its source is public at github.com/ultralytics/yolov5. The weights themselves have not been separately licensed.