README.md

metadata

title: MedGRPO Demo — Medical Video Understanding
emoji: 🏥
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 5.50.0
app_file: app.py
pinned: false
license: apache-2.0
hardware: zero-a10g

MedGRPO — Medical Video Understanding

Interactive demo for uAI-NEXUS-MedVLM-1.0c-4B-SFT (base: Qwen3.5-VL-4B) — part of the uAI-NEXUS-MedVLM 1.0 family, fine-tuned on MedVidBench for medical video question answering. The publicly released family member is uAI-NEXUS-MedVLM-1.0a-7B-RL.

Tasks

Task	Description
Temporal Action Localization	Identify when surgical actions occur (start–end times)
Spatiotemporal Grounding	Locate instruments/anatomy in space and time
Dense Captioning	Time-stamped descriptions of action segments
Next Action Prediction	Predict the next procedural step
Video Summary	Summarize the surgical procedure
Region Caption	Describe activity of a specific instrument/region
CVS Assessment	Score Critical View of Safety criteria
Skill Assessment	Rate surgical skill (1–5 scale)

Data Sources

Examples drawn from 8 medical video datasets: AVOS, Cholec80-CVS, CholecT50, CholecTrack20, CoPESD, EgoSurgery, NurViD, and JIGSAWS.

Links

• Paper
• Project Page
• Model (uAI-NEXUS-MedVLM-1.0a-7B-RL)
• Inference Code
• MedVidBench
• MedVidBench Leaderboard