JEV-Contrast

Dataset Description:

JEV-Contrast is a multiple-choice visual question answering dataset. Each item has one image, one question and one correct answer, and offers two candidate sets. Every option in a set differs from the correct answer in one detail, so each set requires checking a different region of the image. These evidence regions are annotated per set.

Number of options. A set has between 3 and 10 options (the maximum of the Jev Decision Index vision board). In validation and test, both sets of an item have the same number of options, and each value of n has the same number of items within each question type, so the option count is spread evenly instead of piling up at the maximum. In validation and test every wrong option was judged by a vision-language model to be clearly contradicted by the image (train used a faster, less strict check). An n = 3 item has the same options as in v0.3.x.

This dataset is for research and development only.

Dataset Creation Date:

2026-10-16

Version:

0.4.1

License/Terms of Use:

The annotations created for this dataset (questions, candidate sets, evidence regions, tiers, splits) are released under CC BY 4.0. Images and the source annotations they derive from remain subject to their original licences and terms:

source	licence / terms
Visual Genome images and annotations	CC BY 4.0
GQA questions and scene graphs	CC BY 4.0
COCO images	Flickr terms of use (per-image licences)
LVIS v1 annotations	CC BY 4.0
VAW annotations	see the VAW repository

Users are responsible for complying with these terms.

Intended Usage:

Evaluating and training vision-language models on candidate-dependent visual evidence (multiple-choice VQA with per-candidate-set evidence regions).

Dataset Characterization

Data Collection Method

• Hybrid: Automated, Synthetic. Images and object annotations come from public datasets; questions and options are generated from scene graphs and annotations, with model-written candidate options.

Labeling Method

• Hybrid: Automated, Synthetic. Labels come from the source annotations; items are filtered by model-based quality checks. The validation and test splits are also verified by a reasoning judge and a three-model panel.

Dataset Format

Parquet with embedded images (HF Image feature), text fields and structured evidence boxes. Light JSONL annotations without images are in annotations/.

Usage

from datasets import load_dataset
ds = load_dataset("nvidia/JEV-Contrast")
ex = ds["test"][0]
ex["image"]                                   # PIL image
ex["set1_options"], ex["set1_answer_index"]   # options (already shuffled) and the position of the correct one
ex["set1_evidence"]                           # evidence regions for set 1: [{"label", "role", "box_xyxy"}]

Scoring. Options have different counts, so report chance-corrected accuracy per item (acc - 1/n) / (1 - 1/n) averaged over items, as the Jev Decision Index does.

Training. In train, set*_options hold all options available for the set (up to 10, shuffled). Sample the number of options n and which wrong options to keep at training time (always keep the correct one), and mask the answer distribution to the n letters shown. Train sets can have different n; this works as augmentation.

Fields

field	description
id	item id
image	the image
width, height	coordinate frame of the evidence boxes
source_dataset, source_split, source_id	provenance: source dataset (GQA, LVIS v1, VAW), its split, and the original record id (GQA question id for items built from GQA questions, otherwise the image id). Images reserved for our validation/test may come from a source's train split; they never appear in our train split.
image_source, image_id	image provenance for deduplication (visual_genome, coco); Visual Genome photos that are also COCO photos are listed under COCO
question, answer	question and correct answer
set1_options, set2_options	the two candidate sets, 3 to 10 options each, in shuffled order
set1_answer_index, set2_answer_index	position of the correct option in each set
set1_evidence, set2_evidence	evidence regions needed to rule out the other options of that set: label, role (answer or option), box_xyxy
tier	S / R: evidence regions verified as necessary; W (train only): answer verified, evidence-region necessity not guaranteed

Dataset Quantification

split	records
train	29,225
validation	430
test	1,951

Options per set (items with that number of options in the first set; train lists the maximum available):

split	3	4	5	6	7	8	9	10
train	496	505	501	557	920	1355	12854	12037
validation	58	58	58	58	58	58	58	24
test	257	257	257	257	257	257	257	152

Validation and test are balanced per option count (counts for 10 are lower because one question type has at most 9 options).

Total storage: about 4.2 GB.

Preview

image (red / blue = evidence regions of set 1 / set 2)	question	options	set 1 options	set 2 options
	What is on the rocks?	3	vine towel toilet ✓	toilet ✓ post towel
	What is the statue on?	4	pillars headboard nightstand ✓ bench	nightstand ✓ headboard couch dresser
	What is behind the person?	5	A blender behind a woman ✓ Shelves behind a woman A microwave behind a woman A door behind a woman A mirror behind a woman	A blender behind a woman ✓ A blender behind a boy A blender behind a mannequin A blender behind a cupboard A blender behind a man
	Which description fits the image?	6	a bow and a hat an umbrella and a hat a mallet and a hat an ironing board and a hat a kite and a hat ✓ a chopping board and a hat	a kite and a balloon a kite and a hat ✓ a kite and a bow a kite and a backpack a kite and an ironing board a kite and an umbrella
	Which description fits the image?	7	a telephone and a coffee maker a spice rack and a coffee maker ✓ a cutting board and a coffee maker a strainer and a coffee maker a fruit bowl and a coffee maker a chap and a coffee maker a plant pot and a coffee maker	a spice rack and an air‑fryer a spice rack and a juicer a spice rack and a telephone a spice rack and a coffee maker ✓ a spice rack and an electric‑grill a spice rack and a strainer a spice rack and a chap
	What do the bird and the feeder look like?	9	a beige bird and a green feeder a tan bird and a green feeder a white bird and a green feeder a black bird and a green feeder a silver bird and a green feeder a red bird and a green feeder an orange bird and a green feeder a gray bird and a green feeder a blue bird and a green feeder ✓	a blue bird and a purple feeder a blue bird and a beige feeder a blue bird and a green feeder ✓ a blue bird and a pink feeder a blue bird and a tan feeder a blue bird and a silver feeder a blue bird and a gray feeder a blue bird and a yellow feeder a blue bird and a gold feeder
	What is behind the sleeping animal?	10	A pillow behind a cat ✓ A shelf behind a cat A mirror behind a cat A box behind a cat A basket behind a cat A blanket behind a cat A toy behind a cat A clock behind a cat A newspaper behind a cat A laptop behind a cat	A pillow behind a baby A pillow behind a teddy A pillow behind a blanket A pillow behind a book A pillow behind a remote A pillow behind a plant A pillow behind a tissue A pillow behind a toy A pillow behind a shoe A pillow behind a cat ✓

Reference(s):

Source datasets: Visual Genome, GQA, COCO, LVIS, VAW.

Ethical Considerations:

Images come from public photo collections and may show people. Items were generated and filtered automatically and spot-checked by models, not fully verified by humans; some labels or evidence boxes may be wrong. Users should check that the dataset meets their requirements and the applicable terms of the source datasets.