kianasun/MARVEL

Dataset Details

Dataset Description

MARVEL is a new comprehensive benchmark dataset that evaluates multi-modal large language models' abstract reasoning abilities in six patterns across five different task configurations, revealing significant performance gaps between humans and SoTA MLLMs.

Dataset Sources [optional]

• Repository: https://github.com/1171-jpg/MARVEL_AVR
• Paper [optional]: https://arxiv.org/abs/2404.13591
• Demo [optional]: https://marvel770.github.io/

Uses

Evaluations for multi-modal large language models' abstract reasoning abilities.

Dataset Structure

The directory images keeps all images, and the file marvel_labels.jsonl provides annotations and explanations for all questions.

Fields

• id is of ID of the question
• pattern is the high-level pattern category of the question
• task_configuration is the task configuration of the question
• avr_question is the text of the AVR question
• answer is the answer to the AVR question
• explanation is the textual reasoning process to answer the question
• f_perception_question is the fine-grained perception question
• f_perception_answer is the answer to the fine-grained perception question
• f_perception_distractor is the distractor of the fine-grained perception question
• c_perception_question_tuple is a list of coarse-grained perception questions
• c_perception_answer_tuple is a list of answers to the coarse-grained perception questions
• file is the path to the image of the question

Citation [optional]

BibTeX:

@article{jiang2024marvel,
  title={MARVEL: Multidimensional Abstraction and Reasoning through Visual Evaluation and Learning},
  author={Jiang, Yifan and Zhang, Jiarui and Sun, Kexuan and Sourati, Zhivar and Ahrabian, Kian and Ma, Kaixin and Ilievski, Filip and Pujara, Jay},
  journal={arXiv preprint arXiv:2404.13591},
  year={2024}
}