Extend-AI/PoliTax-Split

PolitTax: Document Splitting Benchmark

A benchmark for evaluating document splitting (boundary detection and classification) on U.S. presidential tax returns.

Overview

Metric	Value
PDF files	30
Total pages	3,311
Annotated sub-documents	1,152
Classification types used	40
Subjects	Biden, Harris, Obama, Trump
Tax years	2009–2023

Task

Given a multi-page PDF containing one or more IRS tax forms, schedules, and supplementary documents, the task is to:

1. Segment the document into sub-documents by identifying start and end page boundaries
2. Classify each sub-document by its form type (e.g., "Form 1040", "Schedule A (Form 1040)", "supplement")

Sub-document Type Distribution (top 10)

Type	Count
supplement	573
misc_form	240
Schedule C (Form 1040)	41
Form 4562	40
Schedule K-1 (Form 1065)	33
Schedule E (Form 1040)	14
Form 1065	14
Schedule M-3 (Form 1120/1065)	14
Form 8949	13
Form 4797	13

Files

• annotations.jsonl — One JSON object per PDF with ground-truth sub-document boundaries
• metadata.csv — File-level metadata (president, tax year, page count, etc.)
• taxonomy.json — Full classification schema with 40+ type definitions and splitting instructions
• pdfs/ — Original PDF files organized by subject

Annotation Format

Each line in annotations.jsonl contains:

{
  "file_name": "obama_2009_complete_return.pdf",
  "page_count": 59,
  "sub_documents": [
    {
      "type": "Form 1040",
      "start_page": 1,
      "end_page": 2,
      "identifier": "Barack H. & Michelle L. Obama"
    }
  ]
}

• type: The IRS form or document type. See taxonomy.json for the full list with descriptions.
• start_page / end_page: 1-indexed, inclusive page range.
• identifier: A human-readable label distinguishing instances of the same form type (e.g., taxpayer name, entity name).

Source

All PDFs are publicly released U.S. government records sourced from:

• Obama White House Archives
• The American Presidency Project
• Tax Notes

License

The source documents are U.S. government works in the public domain. The annotations in this dataset are released under CC0 1.0.