mJev-Compositional-VQA

English | 简体中文

mJev-Compositional-VQA is a human-reviewed Chinese visual question answering dataset for mJev-style candidate-based evaluation containing 193 questions grounded in 100 images. It evaluates compositional visual grounding: selecting a target by combining visible attributes, spatial relations and reference objects. The release contains natural photographs from COCO, Open Images and Places365, paired with structured Jev questions and variable-size candidate sets.

Intended use

The dataset uses the Jev question–candidate–label structure and supports exploratory evaluation of Jev/mJev and other vision-language models on multi-cue visual understanding, including:

• Identifying a target using multiple visible conditions.
• Resolving spatial relationships with one or more reference objects.
• Distinguishing similar objects through attribute and relation combinations.

Each question has a single reference answer and three or four candidates. Images may have multiple questions, enabling both question-level and image-level analysis. This compact evaluation set focuses on compositional grounding; it does not assign difficulty scores based on model performance.

Dataset composition

Image source	Images	Questions
COCO	45	87
Open Images	35	66
Places365	20	40
Total	100	193

Questions and candidate text are in Chinese. There are 169 four-choice and 24 three-choice questions; 93 images have two questions and seven images have one.

Data format

data/pilot.parquet contains one row per question. Images are embedded for 189 questions; four questions reference two external Flickr images. For those rows, image and image_path are null, and external_image_url provides the source page. The pilot split can be loaded with Hugging Face Datasets:

Install datasets and Pillow, and authenticate with an account that has repository access when loading a private copy.

pip install datasets Pillow

from datasets import load_dataset

ds = load_dataset("Immortal-Zhang/mJev-Compositional-VQA", split="pilot")
row = ds[0]
candidates = dict(zip(row["candidate_labels"], row["candidates"]))

Field	Description
id, image_id	Stable question and image identifiers
image, image_path	Embedded image and relative image path; null for link-only rows
media_delivery, external_image_url	embedded or external_link; source page URL for link-only rows
question	Chinese question text
candidate_labels, candidates	Aligned option labels and answer texts
label	Reference answer label
question_type	Descriptive question category
source_dataset, source_id, source_url, media_license	Image provenance and source license information

jev.jsonl provides the same questions in nested Jev format. The image value is either a path relative to the dataset root or an external Flickr page URL:

{"id":"<question-id>","image":"images/<source>/<file>.jpg","question":{"instructions":"<question>","criteria":{"A":"<candidate>","B":"<candidate>","C":"<candidate>"}},"label":"A"}

The 98 distributed images are also available under images/. The other two images are external references only; no image bytes or automatic downloads are included. media_manifest.jsonl provides per-image attribution, transformation notes and hashes. stats.json and SHA256SUMS provide statistics and integrity information.

Annotation quality and evaluation

The dataset provides 193 structured Jev multiple-choice questions, each pairing a visual question with semantically aligned candidates and a reference answer label. Its coverage emphasizes multi-condition grounding, reference-based spatial relations, and attribute composition. All 193 questions have been individually reviewed by a human against the images, including the question, every candidate and the reference answer.

For evaluation, provide only the image, question and candidates. Exclude reference answers, annotation evidence and source metadata from model inputs. Report accuracy against the human-reviewed reference labels and group related questions by image when splitting data or estimating uncertainty. These 100 images are a subset of an earlier 5,000-image collection and should not be treated as independent of that collection.

Licenses

Source images retain their original rights and license conditions; the collection is not uniformly licensed under Apache-2.0. Per-image metadata and unresolved attribution or permission details are documented in THIRD_PARTY_LICENSES.md. See LICENSE.md for the collection's rights statement.

Related project

mJev provides candidate-based multimodal inference. This dataset supplies the image, question, candidates and reference label needed for image-based evaluation.

References

• COCO
• Open Images
• Places365