SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

Granite-Docling-258M - A lightweight vision-language model released by IBM

Author: AI小集 · Original source: AI工具集

What is Granite-Docling-258M

Granite-Docling-258M is a lightweight vision-language model released by IBM, designed for efficient document conversion. The model can convert documents into machine-readable formats while fully preserving elements such as layout, tables, and formulas. The model contains only 258M parameters, delivers excellent performance, is highly cost-effective, and supports multilingual processing (including Arabic, Chinese, and Japanese). The model uses the DocTags format to precisely describe document structure and avoid information loss. Granite-Docling-258M can seamlessly integrate with the Docling library, providing powerful customization and error handling capabilities. It is suitable for enterprise-level document processing and is a powerful tool in the field of document processing.

Granite-Docling-258M

Main features of Granite-Docling-258M

  • Precise document parsing: The model can accurately recognize and parse various elements in documents, such as text, tables, formulas, and charts, providing a clear and accurate data foundation for subsequent processing.
  • Structure-preserving conversion: When converting documents into electronic formats, it fully preserves the layout and structure of the original document, ensuring that the converted document is highly consistent with the original and convenient for reading and further editing.
  • Multimodal input support: It supports both image and text input, and can process documents in various forms such as scanned documents, handwritten notes, and electronic documents, broadening the scope of application.
  • Multilingual document processing: It has multilingual processing capabilities and can handle documents in different languages, providing convenience for document processing in multinational enterprises and multilingual environments.
  • Efficient data extraction: It supports quickly extracting key information and structured data from documents, improving work efficiency and reducing manual processing time.
  • Flexible output formats: It supports converting documents into a variety of common formats, such as Markdown, HTML, and JSON, making it convenient for users to perform subsequent processing and applications according to their needs.
  • Powerful customization capabilities: Integrated with the Docling library, users can customize document processing workflows according to specific needs, enabling personalized document conversion and analysis functions.
  • Enterprise-level stability: After optimization, the model is more stable when processing documents, reducing the occurrence of errors and exceptions, and is suitable for large-scale application in enterprise-level environments.

Technical principles of Granite-Docling-258M

  • Model architecture：
    • Vision encoder: It uses siglip2-base-patch16-512 as the vision encoder, which can efficiently process image input and extract visual features from documents.
    • Vision-language connector: Based on a pixel shuffle projector, it connects visual features with the language model to achieve fusion of visual and language information.
    • Language model: Based on the Granite 165M language model, it can process and generate natural language text, ensuring accurate conversion of document content.
  • DocTags format: DocTags is a universal markup language that can precisely describe various elements in documents (such as charts, tables, formulas, etc.) as well as their contextual relationships and positions. : The DocTags format optimizes LLM readability, enabling the model's document output to be directly converted into formats such as Markdown, HTML, or JSON, facilitating subsequent processing and applications.
  • Training data: Training data includes public datasets and internal synthetic datasets, such as SynthCodeNet (code snippets), SynthFormulaNet (mathematical formulas), SynthChartNet (charts), and DoclingMatix (real document pages). Through high-quality annotated data, the model can better learn the structure and content of documents, improving the accuracy and stability of conversion.

Granite-Docling-258M Project Address

  • Project Official Website：https://www.ibm.com/new/announcements/granite-docling-end-to-end-document-conversion
  • HuggingFace Model Repository：https://huggingface.co/ibm-granite/granite-docling-258M
  • Online Experience Demo：https://huggingface.co/spaces/ibm-granite/granite-docling-258m-demo

Application Scenarios of Granite-Docling-258M

  • Enterprise Document Management: The model can quickly digitize paper documents, facilitating storage and retrieval, and improving work efficiency.
  • Academic Research: The model can efficiently process large amounts of literature, helping researchers quickly obtain and analyze materials.
  • Government Archive Digitization: Used for precise conversion of historical archives, ensuring information integrity and facilitating long-term preservation and querying.
  • Education: Teachers can quickly organize teaching materials, and students can conveniently access electronic learning materials.
  • Multilingual Document Processing: Multinational enterprises can process multilingual documents, breaking down language barriers and promoting international communication.