Infinity-Doc-55K

Paper | Github | 💬 Web Demo

Overview

Infinity-Doc-55K is a high-quality diverse full-text parsing dataset, comprising 55,066 real-world and synthetic scanned documents. The dataset features rich layout variations and comprehensive structural annotations, enabling robust training of document parsing models. Additionally, this dataset encompasses a broad spectrum of document types, including financial reports, medical reports, academic reports, books, magazines, web pages, and synthetic documents.

Data Construction Pipeline

To construct a comprehensive dataset for document parsing, we integrate both real-world and synthetic data generation pipelines. Our real-world data pipeline collects diverse scanned documents from various practical domains (such as financial reports, medical records, and academic papers), employing a multi-expert strategy with cross-validation to generate reliable pseudo-ground-truth annotations for structural elements like text, tables, and formulas. Complementing this, our synthetic data pipeline programmatically creates a wide array of documents by injecting content from sources like Wikipedia into predefined HTML layouts, rendering them into scanned formats, and extracting precise ground-truth annotations directly from the original HTML. This dual approach yields a rich, diverse, and cost-effective dataset with accurate and well-aligned supervision, effectively overcoming common issues of imprecise or inconsistent labeling found in other datasets and enabling robust training for end-to-end document parsing models.

Data Statistics

Due to company policy constraints, we are currently open-sourcing only a portion of the training data. The statistics are as follows:

Document Type	Samples Number	BBox and Reading Order
Synthetic Documents	6546	✓
Financial Reports	2000
Medical Reports	2000
Academic Papers	5979
Books	1749
Magazines	1688
Web Pages	2000
All	21962

The full dataset will be released in the future. We appreciate your understanding.

Data Format

Full-page document parsing format:

[
  {
    "images": ["path/to/image"],
    "conversations": [
      {
        "from": "human",
        "value": "prompt"
      },
      {
        "from": "gpt",
        "value": "ground_truth"
      }
    ],
    "attributes": {
      "document_type": "document_type",
      "task": "task"
    }
  },
  ...
]

Layout analysis and reading order format:

[
  {
    "images": ["path/to/image"],
    "form": [
      {
        "category": "The category is in [\"title\", \"plain_text\", \"table\", \"table_caption\", \"figure\", \"figure_caption\", \"header\", \"footer\"]"
        "bbox": [x1, y1, x2, y2],
        "text": "content of the element",
        "html": "table html"
      },
      ...
    ]
  },
  ...
]

Note that the elements in 'form' are sorted in reading order.

License

This dataset is licensed under cc-by-nc-sa-4.0.