EgoF0102/cholec-scene-graph-val

Cholec Scene-Graph Validation Benchmark

This is a self-contained, 2,298-example multiple-choice surgical VQA validation benchmark derived from CholecT45 test videos 74, 75, 78, 79, and 80. It has six subsets: L1 (504), L2 (504), L3 (390), L4 (300), CVS (300), and DYN (300).

Dataset structure

The dataset is Parquet-sharded for Hub preview and streaming. image contains the primary PNG frame as bytes; image_after contains the second frame for DYN examples and is null otherwise. No external absolute paths are required to view these images.

Field	Description
id	Stable benchmark identifier
task, difficulty	Task family and level
image, image_after	Embedded image(s); the latter is used only for dynamic examples
question, choices, answer_idx, answer	Multiple-choice VQA item and gold label
evidence	JSON-encoded source annotation evidence
video, frame_ids, source_image_paths	Reproducibility metadata with relative paths only

Loading

from datasets import Image, load_dataset

ds = load_dataset("parquet", data_files="data/*.parquet", split="validation")
ds = ds.cast_column("image", Image()).cast_column("image_after", Image())
example = ds[0]
example["image"].show()

Once published to the Hub, replace the first two arguments with the repository id:

ds = load_dataset("EgoF0102/cholec-scene-graph-val", split="validation")

Source data and redistribution

The items are generated from CholecT45, Cholec80 1-fps phase labels, and CVS annotations. Check the terms of all source datasets before publishing or redistributing this derivative dataset, particularly the embedded frames.

Generation

The accompanying source_val.jsonl and generate_cholec_benchmark.py preserve the original record structure and reproduce the annotations when supplied with authorized source data. Generation seed: 42.