ML-FIGS 📚📊

This dataset comprises a collection of 4,256 figures and corresponding metadata extracted from 43 different machine learning books. Each directory represents one book and contains a subdirectory with images and a JSON file holding metadata for each figure. The dataset is organized hierarchically as follows:

ml-figs/
├── Book_1/
│   ├── image/
│   │   ├── Figure1.png
│   │   ├── Figure2.png
│   │   └── ...
│   └── Book_1.json
├── Book_2/
│   ├── image/
│   │   ├── Figure1.png
│   │   ├── Figure2.png
│   │   └── ...
│   └── Book_2.json
├── ...
│
├── mlfigs_train.json
│
└── mlfigs_test.json

Each JSON file in the dataset represents the metadata for all the figures within a particular book. A typical entry for a figure in the JSON file includes the following attributes:

• caption: Describes the content of the figure. For example, "Figure 25: Gradient Descent (4/4)".
• captionBoundary: The bounding box of the caption text within the page, represented as a dictionary:
  • x1, x2: Horizontal boundaries of the caption.
  • y1, y2: Vertical boundaries of the caption.
• figType: The type of figure (usually "Figure", "Table", or other structural elements).
• imageText: A list of any text recognized within the figure image.
• name: The unique identifier for each figure within the book (e.g., "25").
• page: The page number on which the figure appears.
• regionBoundary: The bounding box of the entire figure within the page, defined as:
  • x1, x2: Horizontal boundaries of the figure.
  • y1, y2: Vertical boundaries of the figure.
• renderDpi: The DPI (dots per inch) resolution used when rendering the image.
• renderURL: The path to the corresponding figure image file within the image/ directory.
• ocr: OCR (Optical Character Recognition) data, capturing any text detected within the figure.
• text : A list of strings, representing the recognized text within the figure. These are typically individual words or symbols extracted by the OCR system.
• left : Horizontal coordinates (in pixels) representing the left edge of each recognized text element.
• top : Vertical coordinates (in pixels) representing the top edge of each recognized text element.
• width : Width of each recognized text element, providing the horizontal span of the text box.
• height : Height of each recognized text element.
• conf : Each recognized text element is assigned a confidence score ranging from 0 to 100, indicating the OCR system's confidence in its recognition.

📕 Citation

@misc{salamat_kuantaiuly_2025,
    author       = { Salamat Kuantaiuly },
    title        = { ml-figs (Revision 8695d1b) },
    year         = 2025,
    url          = { https://huggingface.co/datasets/salamnocap/ml-figs },
    doi          = { 10.57967/hf/5251 },
    publisher    = { Hugging Face }
}