prg-unibe/dodis-historical-documents

Dataset Overview

The Swiss Historical Archive of DOdis Works (SHADOW) is a large-scale benchmark dataset derived from the resources of Dodis, an independent research center dedicated to the study of Swiss foreign policy and Switzerland’s international relations. Dodis has curated and published nearly 50,000 key documents that document the administrative practices and decision-making processes of the Swiss federal administration.

The underlying corpus consists of notes, letters, minutes, and protocols preserved in both handwritten and typewritten form. All primary documents are released in PDF/A format and have not been subjected to optical character recognition (OCR). For approximately 10,000 documents, Dodis additionally provides manually transcribed and annotated text as part of the Diplomatic Documents of Switzerland (DDS) edition, prepared by professional historians.

In addition to the document collection, Dodis maintains a relational database containing rich structured metadata, including information on actors, organizations, thematic tags, and geographical references. Of particular importance are the expert-written summaries authored by historians, which describe the content and historical relevance of each document and serve as high-quality ground truth labels.

SHADOW builds on the combination of unstructured historical documents and extensive structured metadata provided by Dodis. The dataset comprises 43,155 documents along with their associated metadata. Textual content is available through up to three different extraction methods: manual human transcription, OCR output, and raw PDF text extraction. Depending on the document, one or more of these text representations may be available.

Documents are written in one of Switzerland’s four national languages (German, French, Italian, or Romansh) or in English, with a strong historical and administrative emphasis on German and French. The dataset is partitioned into 34,524 training documents, 4,315 validation documents, and 4,316 test documents. The train–validation–test split is stratified by document type and year of publication to preserve both topical and temporal distributions across all subsets.

Changelog

v1.2.0

This release extends the dataset with multilingual and translation-enhanced fields to improve accessibility and downstream NLP usability.

New features:

• Added combined_text, a unified text representation constructed using a combine_first strategy across all available OCR and transcription sources:

  • transcript_html
  • mistral-ocr-md, mistral-ocr-md-de, mistral-ocr-md-fr, mistral-ocr-md-en
  • pymupdf4llm-md, pymupdf4llm-md-de, pymupdf4llm-md-en
  • transkribus-md, transkribus-md-de, transkribus-md-fr, transkribus-md-en, transkribus-md-it

• Added english_combined_text, an English translation of combined_text (or a direct copy if already in English).

• Added english_summary, an English translation of the original summary field (or a direct copy if already in English).

• Added summary_in_text_language, a translation of the summary into the language of combined_text (or a direct copy if already aligned).

Notes:

• The only preprocessing applied is HTML-to-plaintext conversion where applicable.
• No additional cleaning, normalization, or rewriting of source content was performed beyond aggregation and translation.

---

v1.1.0

• Introduced Transkribus PAGEXML-derived markdown representations wherever available.
• Expanded coverage of transcription sources using structured PAGEXML conversions into Markdown format to improve readability and downstream processing.

---

v1.0.0

• Initial release of the dataset.
• Contains the base collection of historical document data and associated metadata.

Intended Use

SHADOW is designed as a high-quality NLP resource for research on diplomatic and historical texts. The availability of expert-written summaries, multilingual content, and extensive metadata makes the dataset particularly well suited for tasks such as text summarization, document classification, information retrieval, and metadata-aware language modeling. More broadly, the dataset supports a wide range of downstream applications that benefit from large-scale, well-curated textual data paired with reliable annotations. To support the broadest possible range of use cases, users are encouraged to apply task-specific preprocessing to the data. While general cleaning has already been performed, the dataset intentionally retains a wide variety of documents and textual artifacts in order to avoid excluding material that may be valuable for different downstream tasks.