IPO-Mine: A Section-Structured Analysis Toolkit and Dataset for Long-Text Multimodal IPO Documents

This article introduces the IPO-Toolkit open-source framework and the IPO-Dataset. The dataset covers over 109,000 IPO filing documents and amendments from 1994 to 2026, including more than 76,000 images. The study reveals that current multimodal models have significant discrepancies with human experts' judgments when processing ultra-long regulatory documents, providing an important benchmark for multimodal reasoning research on financial documents.

IPO filing documents are important disclosures made by private companies when going public, covering key information such as business models and financial status. However, they present challenges like ultra-long length (often exceeding 500,000 tokens), multimodality, and inconsistent structure. Although large models have made significant progress in document understanding, the lack of large-scale standardized datasets and evaluation benchmarks in the IPO field limits model assessment and improvement.

IPO-Toolkit

• Document segmentation: Automatically split lengthy files into standardized sections
• Image extraction: Extract embedded images and charts from PDFs
• Structured output: Generate structured data for reproducible analysis

IPO-Dataset

• Time span: 1994-2026
• Number of documents: Over 109,000 filing documents and amendments
• Number of images: Over 76,000
• Format: Section-structured text + corresponding image data

Evaluation tasks based on IPO-Dataset focus on financial chart quality assessment and misleading content detection. Results show that current state-of-the-art multimodal models have significant discrepancies with human experts' judgments in these tasks, exposing alignment challenges for models when understanding long-text regulatory documents.

IPO-Dataset supports the following research directions:

• Section-level text variation analysis
• Cross-industry comparison of visual and text disclosure practices
• Temporal evolution of IPO document disclosure standards
• Regulatory compliance analysis and corporate response strategy research

The research team has open-sourced resources such as code and datasets under the CC-BY-4.0 license, which helps to:

• Promote reproducible research in financial AI
• Lower the entry barrier for new researchers
• Establish industry standards and best practices
• Drive practical applications of multimodal document understanding technology

Limitations

• Significant alignment gap between models and human experts
• Dataset is mainly based on the U.S. market
• More fine-grained annotations are needed for chart misleading detection

Future Directions

• Improve model training by integrating domain expert knowledge
• Extend the toolkit to other financial document analysis tasks