SUREON

GitHub · Website · arXiv

Surgical video question answering from SUREON: A Benchmark and Vision-Language Model for Surgical Reasoning, accepted at WACV 2027.

Splits

Split	All questions	Passing blind curation	Label provenance
Train	176,591	76,061	Automatic pseudo-labels
Validation	7,039	2,081	Automatic pseudo-labels
Test	340	Not applicable	Surgeon-verified

Splits are disjoint at the full source-video level, including known YouTube aliases. Training and validation are not surgeon-verified. Complementary public perception datasets used in SUREON-VLM are not included.

The test benchmark has 340 questions over 335 distinct clip IDs and ten categories. Score 299 questions across nine categories for accuracy, including temporal ordering. Evaluate the 41 sequence summaries separately. Use question_id for predictions; clip_id is not a unique question identifier.

Train and validation contain the eligible pools before blind curation. Select passes_blind_curation == true for the curated subsets used by the paper. false denotes a recorded rejection; null denotes no recorded decision and is also used for the expert test set, where this flag does not apply. Sequence-summary passthroughs count as passing according to the original pipeline.

Annotation format

The default dataset configuration exposes train, validation, and test directly from typed Parquet shards under data/.

Fields	Meaning
question_id, clip_id, split	Unique question ID, clip/window ID, and split
question_type, task_format	Category and expected response format
question, choices, answer_text	Question, fixed options where applicable, and reference answer
reference_text	Sequence-summary reference; null for other tasks
distractor_pools, baseline_distractors	Training distractor candidates
source_video_id, youtube_id, source_url	Source lecture identifiers and access URL
start_timestamp, end_timestamp, start_seconds, end_seconds, duration_seconds	Exact annotated temporal window
specialty, procedure, label_provenance	Available source metadata and label origin
passes_blind_curation	Nullable recorded curation decision

Validation/test options are fixed. Training choices is empty because options are sampled from the released distractor pools at runtime. Temporal descriptions remain in question, with their reference order in answer_text; temporal ordering and summarization have no choices. Question wording and reference answers are preserved.

Video access and limitations

Original lecture videos are not redistributed and retain their source terms. The annotations provide source URLs/YouTube IDs and exact windows. See the SurgLaVi download instructions.

License

Released under CC BY-NC-SA 4.0: https://creativecommons.org/licenses/by-nc-sa/4.0/

Citation

If you use SUREON, please cite:

@article{perez2026sureon,
  title={SUREON: A Benchmark and Vision-Language Model for Surgical Reasoning},
  author={Perez, Alejandra and Rau, Anita and Mlambo, Busisiwe and White, Lee and
          Brunet, Aina and Verde, Juan and Caicedo, Juan and Lavanchy, Jo{\"e}l L. and
          Schmidt, Adam and Nwoye, Chinedu and Mohareri, Omid and Jamal, Muhammad Abdullah},
  journal={arXiv preprint arXiv:2603.06570},
  year={2026}
}