ContextualBench / README.md
ToughStone's picture
Update README.md
528f8cf verified
metadata
license: apache-2.0
task_categories:
  - image-to-text
  - visual-question-answering
language:
  - en
tags:
  - context-violating
  - visual-language
pretty_name: ContextualBench

Challenging and Enhancing the Reasoning Capacity of Multimodal LLMs in Context-violating Images

Hongxi LiYuyang ChenYayun QiXinxiao Wu
 Beijing Institute of Technology
arXiv 2024
🌎Website (Comming soon) | 🧑‍💻Code | 📄arXiv (Comming soon) | 🏆 Leaderboard

dataset description

Dataset Description

ContextualBench consists of 6(categories) × 12 (instances) = 72 context instances, with each context instances containing 7 context-consistent images and 7 context-violating images. We design 4 visual reasoning tasks and collect human annotations.

Dataset Construction

Image Generation

We collect context instances from external database and formulate them as key-value pairs(as shown in ‘Constraint’ field of Dataset Viewer), then edit a subset of these pairs to make context-violating images. We generate images using text-to-image models. Each image is generated iteratively until the following three conditions are met:

  1. The image conforms to the given constraints.
  2. There are no visual illusions in the image.
  3. The image is free of ambiguity and potential offensiveness.

Task Design

We design 3 visual reasoning tasks for all images in ContextualBench: visual question answer, image caption, and image identification. An additional image explanation task is designed for context-violating images.

Image Annotation

For image caption, image identification, and image explanation, we collect 5 annotations from different human. For visual question answer, we generate Q-A pairs based on image caption following Q2 pipeline.

Licensing Information

  1. Purpose: The dataset was primarily designed for use as a test set.
  2. Commercial Use: Commercially, the dataset may be used as a test set, but it's prohibited to use it as a training set.
  3. Rights on Images: All rights to the images within the dataset are retained by the ContextualBench authors.

Citation Information

@article{hongxi2024challenging,
  title={Challenging and Enhancing the Reasoning Capacity of Multimodal LLMs in Context-violating Images},
  author={},
  journal={},
  year={2024}
}