Surgical Report Generation - Eval Task
Private dataset. See the accompanying code package (build_surgery_report_eval.py + README,
sent separately) for the full task description, scoring guidance, and JSONL schema.
Quick summary: given a surgical video clip where a surgeon flagged certain moments (start of a
phase, start of a (finer-grained) step, or an adverse event), generate a report describing what
happened at those moments. Two formats: blind (only timestamps given, model must identify what
happened) and open_book (facts given, model must summarize). Source: MultiBypass140.
Adverse events cover 4 real types with real severity (bleeding, mechanical injury, thermal injury, ischemic injury - not just bleeding), parsed from the raw per-frame annotation. Every short/medium clip is guaranteed to contain at least one phase-or-step moment AND at least one adverse-event moment (verified by brute-force check before cutting, not assumed).
Three tiers
clips_short/+surgery_report_eval_short.jsonl- 687 clips (1,374 rows), avg ~5.5 min, anchored on real adverse-event clusters with a guaranteed nearby phase/step landmark.clips_medium/+surgery_report_eval_medium.jsonl- 505 clips (1,010 rows), avg ~22 min, same guarantee, wider budget.full_videos/+surgery_report_eval_full_video.jsonl- all 138 qualifying source videos, unedited (37 min - ~3 hours), every real phase/step/adverse-event moment in the whole case - a dedicated long-video-handling stress test, deliberately kept as the full set.
All videos re-encoded to 1fps / 480px-wide / mpeg4, no audio, purely to keep file size down.
- Downloads last month
- 5