BERD@NFDI - NFDI for Business, Economic and Related Data (BERD 2.0)

The project investigates the use of modern Vision Large Language Models (Vision-LLMs) to extract and structure information from historical economic and financial sources. Building on prior collaboration between SAFE and BERD@NFDI, it explores how multimodal AI models can enhance existing OCR-based workflows, particularly for complex historical documents with heterogeneous layouts and tables.
Different Vision-LLMs will be systematically evaluated and compared with established OCR approaches in terms of recognition quality and runtime performance. A particular focus will be on locally deployable models, fine-tuning strategies, and the development of suitable training datasets. The methods will be tested on various historical sources, including materials being made accessible for research within the long-term project on historical financial and company data.
The project aims to develop transferable and reproducible workflows for converting historical documents into structured research data. Training data, models, methodological insights, and workflows are intended to be made available to the research community in accordance with Open Science principles.

Project partner: UB Mannheim
Project team: Dennis Gram (PI, SAFE), Nicolas Zöllner (SAFE), Jan, Kamlah (UB Mannheim)
Project duration: 2 years