yuheng-chen/Unified_Document_Understanding_Dataset

Dataset Card for Read-Parsing-Describe: Unified Scientific Document Understanding

Read-Parsing-Describe (Unified Scientific Document Understanding) is a pioneering multimodal benchmark designed to train and evaluate models on the complex structures of scientific documents. Unlike traditional document datasets that treat visual elements merely as isolated layout blocks, RPD transforms document parsing into an accessibility-driven, cross-modal grounding task.

🌟 Key Highlights & Innovations

• Strict Spatial-to-Semantic Binding: RPD enforces a rigorous agreement between physical bounding boxes, semantic labels, logical reading order, and visual descriptions (alt-text) within a single, auditable page parse.
• Accessibility-Driven Alt-Text Generation: Complex graphics (charts, tables, equations) in scientific papers contain essential knowledge. By mandating context-aware alt-text grounded to exact spatial coordinates, this dataset helps mitigate VLM hallucinations and makes document reconstruction fully accessible to screen readers.
• Complex Multimodal Interactions: Curated specifically to challenge models with the hardest parsing scenarios, including multi-column prose, spanning figures, dense tables, and intertwined mathematical equations.

This unified schema provides a robust foundation for training Parameter-Efficient Fine-Tuning (PEFT) adapters or full Vision-Language Models (VLMs) to master complex layout comprehension and precise cross-modal alignment.

Dataset Structure

The dataset is built using the ImageFolder builder and relies on metadata.jsonl files to map images to their respective annotations.

Data Splits

The dataset is divided into three standard splits:

• train: For model training.
• validation: For hyperparameter tuning and early stopping.
• test: For final model evaluation.

Data Fields

Each instance in the dataset represents a single document page and contains the following fields:

• image (Image): The PIL Image object containing the document page.
• file_name (String): The base name of the image file (e.g., page_001.png).
• source_folder (String): The name of the original folder from which the image was sourced.
• document_id (String): A unique identifier for the document or paper the page belongs to.
• source_image_path (String): The original absolute or relative path to the image file prior to uploading.
• source_json_path (String): The original path to the source JSON file containing the raw annotations.
• annotation_json (String): A serialized JSON string (or dictionary) containing the detailed annotations for the image (e.g., bounding boxes, semantic labels, reading order, and alt-text descriptions).

Usage

You can easily load and iterate through the dataset using the Hugging Face datasets library.