README.md

metadata

language:
  - en
  - fr
  - es
license: mit
size_categories:
  - n<1K
task_categories:
  - object-detection
  - image-segmentation
tags:
  - pdf
  - document-layout-analysis
  - data-extraction
configs:
  - config_name: annotations
    data_files:
      - split: unhcr
        path: annotations/unhcr/*.json
      - split: prwp
        path: annotations/prwp/*.json
      - split: refugee
        path: annotations/refugee/*.json
  - config_name: metadata
    data_files:
      - split: unhcr
        path: metadata/unhcr/*.json
      - split: prwp
        path: metadata/prwp/*.json
      - split: refugee
        path: metadata/refugee/*.json
  - config_name: documents
    data_files:
      - split: unhcr
        path: documents/unhcr/*.pdf
      - split: prwp
        path: documents/prwp/*.pdf
      - split: refugee
        path: documents/refugee/*.pdf
  - config_name: snapshots
    data_files:
      - split: unhcr
        path: snapshots/unhcr/*.png
      - split: prwp
        path: snapshots/prwp/*.png
      - split: refugee
        path: snapshots/refugee/*.png

Dataset card for data-snapshot

This dataset was introduced in the paper Benchmarking Open-Source Layout Detection Models for Data Snapshot Extraction from Institutional Documents.

The source code for the benchmark and dataset extraction is available on GitHub: worldbank/ai4data.

Dataset summary

The data-snapshot dataset is an annotated corpus designed for the evaluation and development of models for extracting data snapshots from PDF documents. A data snapshot is defined as a figure or table containing structured or semi-structured information intended for analytical interpretation or operational reuse.

The benchmark spans humanitarian reports, World Bank policy research working papers, and project appraisal documents, and includes annotations for figures and tables that contain reusable analytical information.

Dataset structure

The repository is organized as follows:

ai4data/data-snapshot/
├── annotations/<source>/*.json  # Contains annotation files per document
├── documents/<source>/*.pdf     # Actual PDFs
├── metadata/<source>/*.json     # Document-level metadata
├── schemas/*.json               # Provides the schema of the annotation and metadata files
├── snapshots/<source>/*.png     # Image files corresponding to the annotations
└── README.md

Subsets

• annotations
  • JSON files that indicate the data snapshots: their object class (Figure / Table) and bounding box locations (in normalized [x1, y1, x2, y2] format, top-left origin)
  • Follows the schema provided in schemas/data-snapshot-eval-v1.3.schema.json
  • Provided on a per-document basis; documents that do not have data snapshots will still have an annotation file present but list of bounding boxes will be empty.
• documents
  • Actual PDF files that were annotated
• metadata
  • Document-level metadata following the World Bank Metadata Standards (Chapter 5 — Documents), schema provided in schemas/metadata_schema.json.
  • Provided on a per-document basis
  • All files across all sources share a uniform schema (same keys at every nesting level)
• snapshots
  • PNG files extracted from the documents and bounding box locations

Sources

• UNHCR
• PRWP
• Refugee

Loading the dataset using HF's datasets library

Annotations

>>> from datasets import load_dataset
>>> annotations = load_dataset("ai4data/data-snapshot", name="annotations", split="unhcr")
>>> annotations[0]  # Inspect the first record
{'label_map': {'1': 'Figure', '2': 'Table'}, 'info': {'schema_version': '1.3', 'type': 'ground_truth', 'created_at': datetime.datetime(2026, 5, 20, 13, 44, 29), 'run_id': 'human-annotation-combined-e3432dce89', 'model': {'name': 'human annotation'}, 'coordinate_system': {'type': 'normalized_xyxy', 'range': [0.0, 1.0], 'origin': 'top_left'}}, 'documents': [{'doc_id': '06072015-baalbek-hermelgovernorateprofile.pdf', 'doc_name': '06072015-baalbek-hermelgovernorateprofile.pdf', 'doc_path': 'pdf_input/06072015-baalbek-hermelgovernorateprofile.pdf'}], 'predictions': [{'page_id': '06072015-baalbek-hermelgovernorateprofile.pdf::p000', 'doc_id': '06072015-baalbek-hermelgovernorateprofile.pdf', 'page_index': 0, 'objects': [{'id': '1d69f693', 'label': 'Figure', 'bbox': [0.029415499554572243, 0.1766403810171256, 0.5954839424856321, 0.7354445202645015], 'score': None}, ...}

Metadata

>>> metadata = load_dataset("ai4data/data-snapshot", name="metadata", split="unhcr")
>>> metadata[0]  # Inspect the first record
{'type': 'document', 'metadata_information': {'title': 'Lebanon: Baalbek-Hermel Governorate Profile (June 2015)', 'idno': '06072015-baalbek-hermelgovernorateprofile', 'producers': [{'name': 'UNHCR', 'abbr': 'UNHCR', 'affiliation': 'UNHCR', 'role': 'Source'}], 'production_date': datetime.datetime(2026, 5, 21, 0, 0), ...}

Documents

>>> docs = load_dataset("ai4data/data-snapshot", data_dir="documents/unhcr")  # Or simply data_dir="documents/" for all files
>>> docs.save_to_disk("path/to/docs_directory")  # Files are saved as an Arrow file

Snapshots

>>> snapshots = load_dataset("ai4data/data-snapshot", data_dir="snapshots/unhcr")  # Or simply data_dir="snapshots/" for all snapshots
>>> snapshots.save_to_disk("path/to/snapshots_directory")  # Files are saved as an Arrow file

Schema

Annotations

The annotation files follow the Data Snapshot Evaluation Format (v1.3). Below is a simplified, human-readable example of the JSON schema with explanatory comments for each field.

Note: You will notice a top-level field called predictions. In the context of this dataset, this is a misnomer because these are actually human-labeled annotations (ground truth). We use the key predictions because we borrow this schema from the project's evaluation codebase, which uses a unified structure for both ground truth and model predictions.

{
  // Canonical mapping of integer IDs to class names
  "label_map": {
    "1": "Figure",
    "2": "Table"
  },
  
  // High-level metadata about the file
  "info": {
    "schema_version": "1.3",
    "type": "ground_truth",  // Indicates these are human annotations
    "created_at": "2026-05-20T13:44:29",
    "run_id": "human-annotation-combined-e3432dce89",
    "model": {
      "name": "human annotation"
    },
    "coordinate_system": {
      "type": "normalized_xyxy",
      "range": [0.0, 1.0],  // Bounding boxes are normalized between 0 and 1
      "origin": "top_left"
    }
  },
  
  // List of documents referenced in this file
  "documents": [
    {
      "doc_id": "1_advocacy_note_mineaction_-_niger_eng.pdf",
      "doc_name": "1_advocacy_note_mineaction_-_niger_eng.pdf",
      "doc_path": "pdf_input/1_advocacy_note_mineaction_-_niger_eng.pdf"
    }
  ],
  
  // Per-page container of objects; these contain the ground truth annotations
  "predictions": [
    {
      "page_id": "1_advocacy_note_mineaction_-_niger_eng.pdf::p001",
      "doc_id": "1_advocacy_note_mineaction_-_niger_eng.pdf",
      "page_index": 0,  // 0-indexed page number
      "objects": [
        {
          "id": "obj_001",
          "label": "Figure",  // Matches a label_map entry
          "bbox": [0.029, 0.177, 0.595, 0.735],  // Normalized [x_min, y_min, x_max, y_max]
          "score": null  // Always null for ground truth
        }
      ]
    }
  ]
}

Metadata

The metadata files follow the World Bank Document Metadata Schema. See schemas/metadata_schema.json for the formal JSON schema definition.

All metadata files across all sources share a uniform schema (same keys at every nesting level, same types) to ensure compatibility with Apache Arrow and HuggingFace streaming.

Top-level fields:

• type
• metadata_information
• document_description
• provenance
• tags
• schematype
• additional - contains source-specific fields (e.g. additional.unhcr_* for UNHCR, additional.wds_* for WDS API-sourced datasets).

Dataset creation

The annotations were produced through human labeling using Label Studio.

Licensing information

MIT License — see LICENSE for details.

Citation information

@misc{dy2026benchmarkingopensourcelayoutdetection,
      title={Benchmarking Open-Source Layout Detection Models for Data Snapshot Extraction from Institutional Documents}, 
      author={AJ Carl P. Dy and Aivin V. Solatorio},
      year={2026},
      eprint={2606.06242},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.06242}, 
}