WebLI-100B - A 100 billion visual language dataset released by Google DeepMind

What is WebLI-100B?

WebLI-100B, released by Google DeepMind, is a massive dataset containing 100 billion image-text pairs for pre-training Visual Language Models (VLMs). An expanded version of the WebLI dataset, WebLI-100B is built upon a vast collection of images and their corresponding titles or page titles gathered from the web as text pairing information. Ten times larger than the previous largest visual language dataset, WebLI-100B leverages massive amounts of data to enhance models' understanding of long-tail concepts, cultural diversity, and multilingual content. Researchers performed only basic data filtering during its construction, preserving as much linguistic and cultural diversity as possible. The emergence of WebLI-100B provides a crucial foundational resource for training more inclusive multimodal models.

Main functions of WebLI-100B

• Support for large-scale pre-trainingWebLI-100B provides 100 billion image-text pairs, offering rich data resources for the pre-training of VLMs and significantly improving the performance of models on various tasks.
• Enhancing cultural diversityIt contains images and text from different cultural backgrounds, helping the model better understand and generate visual and linguistic content related to different cultures.
• Enhance multilingual capabilitiesThe dataset contains text in multiple languages, which helps improve the model's performance on low-resource languages and promotes the development and application of multilingual tasks.
• Support multimodal tasksWebLI-100B data is used in a variety of multimodal tasks, such as image classification, image caption generation, and visual question answering, providing strong support for the development of multimodal models.

WebLI-100B Technical Principles

• Data collection:
  • sourceWebLI-100B's data primarily originates from the internet, collecting images and their corresponding text descriptions (such as image alt text or page titles) through large-scale web crawling.
  • scaleThe dataset contains 100 billion image-text pairs, making it one of the largest visual language datasets to date.
• Data Filtering:
  • Basic FilteringTo ensure data quality and diversity, WebLI-100B only performs basic data filtering, such as removing harmful images and personally identifiable information (PII), in order to preserve as much linguistic and cultural diversity as possible.
  • Quality filtration (optional)The study also explored the use of models such as CLIP for data filtering to improve data quality, but such filtering may reduce the representativeness of certain cultural backgrounds.
• Data processing:
  • Text processingThe image's alt text and the page title are used as paired text, and a multilingual MT5 word segmenter is used for word segmentation to ensure the diversity and consistency of the text data.
  • Image processingAdjust the image to a resolution of 224×224 pixels to meet the input requirements of the model.

WebLI-100B project address

• arXiv technical paper:https://arxiv.org/pdf/2502.07617

Application scenarios of WebLI-100B

• Artificial intelligence researchersUsed for model pre-training, exploring new algorithms, and improving the performance of visual language models.
• engineerDevelop multilingual and cross-cultural applications, such as image description, visual question answering, and content recommendation systems.
• Content creatorsGenerate multilingual image descriptions and tags to enhance the localization and diversity of content.
• Cross-cultural researchers: Analyze images and texts from different cultural backgrounds to study cultural differences.
• Educators and studentsAs a teaching resource, it helps students learn multimodal data processing and analysis.