HumaniBench-MCQ: Task 3 (Multiple-Choice VQA) in MMMU Format

This dataset is a format conversion of Task 3 (task3_Multiple_Choice_VQA) from HumaniBench, restructured to match the MMMU schema so it can be used directly with MMMU-style evaluation pipelines.

No questions, images, or answers were added, removed, or changed. Only the column layout and answer format differ from the original.

---

How to Load

from datasets import load_dataset

ds = load_dataset("vector-institute/HumaniBench-MCQ", split="test")

---

Dataset Structure

1,844 multiple-choice questions, one image each, in a single test split.

Column	Type	Description
id	string	MMMU-style ID: test_<Attribute>_<n>, numbered within each attribute (e.g., test_Gender_1)
question	string	Question text, prefixed with the <image 1> placeholder
options	string	String-formatted list of answer choices, without letter prefixes (e.g., ['Male', 'Female', 'Non-binary', 'Gender-neutral'])
explanation	string	Reasoning supporting the answer (optional; not required for evaluation)
image_1	image	Embedded image data
img_type	string	Image type, always ['Photographs']
answer	string	Single-letter answer (e.g., B)
demographic	string	HumaniBench attribute being tested (e.g., Gender, Age). Used as the categorical field for grouping results (takes the place of MMMU's topic_difficulty)
question_type	string	Always multiple-choice
subfield	string	Left empty
original_id	string	ID from the original HumaniBench Task 3, for traceability

Example

id:               test_Gender_2
question:         <image 1> What gender representation is depicted by the statue of Lady Justice?
options:          ['Male', 'Female', 'Non-binary', 'Gender-neutral']
answer:           B
demographic:      Gender
question_type:    multiple-choice
original_id:      c6caaf72e5

---

Differences from MMMU

• Only image_1 is included, since every question uses a single image. MMMU's image_2 to image_7 columns are omitted.
• MMMU's topic_difficulty column is replaced by demographic, which holds the attribute being tested.
• original_id is an extra column not present in MMMU.

---

Intended Use

• Evaluating multimodal models on human-centric multiple-choice VQA using MMMU-compatible evaluation code
• Comparing model accuracy across social attributes via demographic

For the full benchmark (7 tasks, 32,000+ image–question pairs), see the original HumaniBench dataset.

---

Related Resources

• Original HumaniBench Dataset
• Project Website
• GitHub Repository
• Paper (arXiv)
• MMMU Dataset (format reference)

---

License

This dataset is licensed under the CC BY-NC-SA 4.0 License, the same as the original HumaniBench.

---

🙏 Acknowledgments

Resources used in preparing this research were provided, in part, by the Province of Ontario, the Government of Canada through CIFAR, and companies sponsoring the Vector Institute (vectorinstitute.ai/#partners).

This research was funded by the European Union's Horizon Europe research and innovation programme under the AIXPERT project (Grant Agreement No. 101214389), which aims to develop an agentic, multi-layered, GenAI-powered framework for creating explainable, accountable, and transparent AI systems.

---

📚 Citation

If you use this dataset, please cite the original HumaniBench paper:

@article{raza2026humanibench,
  author = {Raza, Shaina and Narayanan, Aravind and Khazaie, Vahid Reza and Vayani, Ashmal and Radwan, Ahmed Y. and Chettiar, Mukund S. and Singh, Amandeep and Shah, Mubarak and Pandya, Deval},
  title = {HumaniBench: A Human-Centric Framework for Large Multimodal Models Evaluation},
  year = {2026},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  issn = {2157-6904},
  url = {https://doi.org/10.1145/3845999},
  doi = {10.1145/3845999},
  note = {Just Accepted},
  journal = {ACM Trans. Intell. Syst. Technol.},
  month = sep,
}