Sum-of-Checks

Human rubric labels, original laparoscopic images, and original Critical View of Safety (CVS) labels accompanying Sum-of-Checks: Structured Reasoning for Surgical Safety with Large Vision-Language Models.

Subsets

Config / split	Images	Source
endoscapes / few_shot	4	Paper exemplars, in manifest order
endoscapes / dev	49	Remaining corrected Endoscapes annotations
sages / few_shot	6	All v5 examples, in manifest order
sages / dev	54	Remaining SAGES annotations

These are the human-labeled rubric subsets, not the paper's 791-frame evaluation set. The splits are disjoint by image identity and image checksum, with all 113 images retained exactly once. They are not video-disjoint. source_split retains the original val/train provenance. is_few_shot marks both sources; few_shot_order is zero-based prompt order, or -1 for dev. is_paper_few_shot retains its historical Endoscapes-only meaning. SAGES is an additional annotation subset, not an experiment reported in the Endoscapes paper.

The rubric annotator was an ML PhD student trained by a surgeon. These rubric annotations are development data, not expert surgical ground truth. Original source CVS labels are retained separately.

Few-shot examples

The exact four exemplars (CVS patterns 000, 111, 110, 001) are also available as a standalone selection manifest and original JPGs in few_shot_examples/endoscapes/rubrics/filtered/images/. These duplicate only the corresponding Parquet images, not additional examples. The manifest preserves the paper selection order, CVS targets, and all 19 rubric answers. Its saved_image paths resolve from the downloaded repository root.

SAGES retains the six-candidate v5 manifest and its original JPGs. All six enter few_shot in manifest order: patterns 000, 111, 110, 001, 101, 100. None of these six images appear in dev. This retains the full v5 set, not the four-example subsample used in one SAGES experiment; the paper itself reports Endoscapes experiments only. SAGES manifest paths were made repository-relative; all other values and source image bytes are unchanged. The selection version v5 is independent of the rubric specification v3.

from datasets import load_dataset
few_shot = load_dataset("BrachioLab/sum-of-checks", "endoscapes", split="few_shot")
sages_few_shot = load_dataset("BrachioLab/sum-of-checks", "sages", split="few_shot")

For saved image examples and loading in paper order, see the example notebook.

Few-shot Video and Frame Reference

Rows follow the full few-shot manifest order for each source.

Source	Original video filename / ID	Frame ID	Image filename	CVS targets
endoscapes	ID 129	69650	129_69650.jpg	000
endoscapes	ID 154	33500	154_33500.jpg	111
endoscapes	ID 152	41050	152_41050.jpg	110
endoscapes	ID 148	21675	148_21675.jpg	001
sages	24b701a0-f22d-4cba-9871-fab47bd18347.mp4	2400	frame_002400.jpg	000
sages	79b1e745-0fc7-45e0-8159-bcc027cc6db8.mp4	1800	frame_001800.jpg	111
sages	d171f114-19ac-4e30-b24e-ea8015e90bab.mp4	1050	frame_001050.jpg	110
sages	7a208039-f40e-4f03-b73d-555ec9b23119.mp4	900	frame_000900.jpg	001
sages	15c0e57a-97d2-4add-8404-063e3747cd47.mp4	150	frame_000150.jpg	101
sages	3b333490-f68d-4a7c-9e38-14c621da1c85.mp4	1050	frame_001050.jpg	100

SAGES filenames are relative to the original dataset's train/videos/ directory; frame IDs are the original video frame indices used to extract the JPGs. Endoscapes source metadata supplies numeric video IDs, not original video filenames. Its frame IDs above are the original frame numbers encoded in <video_id>_<frame_id>.jpg, not the reindexed frame_id in annotation_coco_vid.json. Original videos are not included in this image-and-label upload.

Fields

• image: original image-file bytes embedded in Parquet; no resizing or re-encoding. SAGES images are the existing JPG video frames used during labeling.
• cvs_labels: original Endoscapes continuous ds values, or majority-vote SAGES criterion labels. cvs_binary_labels thresholds at greater than 0.5.
• cvs_rater_labels: all three original SAGES votes for each criterion, in rater1/rater2/rater3 order; empty for Endoscapes, whose source here supplies continuous ds labels.
• original_cvs_annotation_json: complete source Endoscapes image-annotation object or SAGES frame.csv row, preserving the source fields and values.
• annotation_cvs_labels: CVS labels recorded alongside the manual rubric annotation, separately retained from the original dataset labels.
• rubric_labels: answers to all 19 checks, keyed by item IDs such as C1-1; values are yes, no, or uncertain.
• rubric_version: bundled specification cvs_rubrics_v3. See the rubric JSON for check text, types, and weights.
• annotation_rubric_version: cvs_rubrics_v3 for both Endoscapes and SAGES, following the dataset maintainer's version attribution. This metadata update does not change any rubric answers.
• source_annotation_rubric_version: verbatim metadata from the archived annotation file (cvs_rubrics_v1 for Endoscapes, cvs_rubrics_v3 for SAGES), retained only for source traceability. Legacy Endoscapes filenames are unchanged.
• source_dataset, source_split, source_subset, video_id, frame_id, and annotation source/row identify provenance. Notes and timestamps retain the manual annotation context.
• image_sha256, width, and height support image integrity checks.

The three criteria are: C1, exactly two structures entering the gallbladder; C2, clearance of the hepatocystic triangle; C3, lower-third gallbladder detachment from the liver bed. Sum-of-Checks aggregation uses the v3 weights with yes=1 and no/uncertain=0.

Load

from datasets import load_dataset
endo = load_dataset("BrachioLab/sum-of-checks", "endoscapes", split="dev")
sages = load_dataset("BrachioLab/sum-of-checks", "sages", split="dev")
image = endo[0]["image"]  # PIL image
labels = endo[0]["rubric_labels"]

The rubric JSON is a separate repository file and can be downloaded with huggingface_hub.hf_hub_download. export_manifest.json records image checksums, annotation-source checksums, and the rubric checksum.

Sources and terms

• Endoscapes: CAMMA dataset, CC BY-NC-SA 4.0; see license.
• SAGES: CAMMA SAGES CVS Challenge 2024, CC BY-NC 4.0; see license and dataset paper.
• Rubric annotations and v3 specification: BrachioLab Sum-of-Checks.

Source-specific image/label licenses continue to apply; this export does not replace them with a blanket permissive license. This small, selected annotation set is intended for research and does not establish clinical performance.