← Back to catalog

MedVidBench not scored

Large-scale benchmark for heterogeneous medical and surgical video understanding, comprising 531,850 video-instruction pairs assembled from 8 source datasets (AVOS, CholecT50, CholecTrack20, Cholec80-CVS, CoPESD, EgoSurgery, JIGSAWS, NurViD) and spanning video-, segment-, and frame-level tasks: temporal action localization, spatiotemporal grounding, video summarization, region captioning, and surgical skill / Critical View of Safety (CVS) assessment. Released alongside the MedGRPO model by Northeastern University and United Imaging Intelligence (CVPR 2026). The public Hugging Face distribution provides a ~18 GB archive (103,742 video frames; 6,245 test samples across 8 surgical-video tasks) downloadable as ZIP or via Git LFS. Licensed CC-BY-NC-SA-4.0; individual source datasets retain their own upstream licences and must be complied with separately.

Not auto-scored

This entry has no harvestable repository signal (e.g. a dataset, standard, or hosted service), so the deterministic scorer does not assign a ring.

Signals & lineage

Every value records the source it was harvested from.

FieldValueSource
HuggingFace downloads (monthly)269huggingface
PyPI downloads (monthly)rate_limitedpypi

Change history

Recorded deviations — when a harvested value changed between runs.

WhenFieldWasNowSource
2026-11-20T21:45:51.298495hf_downloads—269huggingface
2026-11-20T21:45:51.242570pypi_downloads——pypi