rbryant19/opscribe-captioner-merged-v1

OpScribe Captioner Combined v1

LoRA adapter for Qwen2.5-VL-72B-Instruct — surgical frame captioning for the OpScribe pipeline.

Training

• Base model: Qwen/Qwen2.5-VL-72B-Instruct
• Dataset: EgoSurgery (9,618 frames) + surgeon-narrated voiceover dataset (3,402 frames, 3x weighted)
• Total training examples: 13,020 train / 1,893 val
• Framework: Custom PyTorch training loop, LoRA rank=16, alpha=32, lr=2e-4
• Epochs: 3 (24h on 4x NVIDIA H200 140GB)
• Best val_loss: 0.1046
• Hardware: CHTC bhaskargpu4000

Usage

from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from peft import PeftModel

base = Qwen2VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2.5-VL-72B-Instruct", device_map="auto"
)
model = PeftModel.from_pretrained(base, "rbryant19/opscribe-captioner-merged-v1")
processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-72B-Instruct")

Framework versions

• PEFT 0.18.1