«Hack Apertus» Liebefeld · Cases
Deciphering Archives with Apertus
1. 1
See the challenge in our Guide
https://app.notion.com/p/hackapertus/University-of-Bern-0a7b4fec112a820a9a1181bc0b5d4d52?source=copy_link
Context and Background
The Swiss Federal Archives (Schweizerisches Bundesarchiv) contain vast historical records that are invaluable for research. We hand-picked three published official open datasets to illustrate a wide spectrum of opportunity for AI-driven inquiry. The preparation work they would involve can be a blueprint for AI-ready data: and exactly what a fully open model like Apertus can now help to automate and extend reproducibly, transparently, at scale.
Problem Description
Transforming raw historical archival data into structured, machine-understandable formats while respecting the original document’s provenance and ethical constraints is an important duty of government archives. Current datasets contain partial preprocessing but lack systemic indexing, NLP-ready formatting, and metadata standardization. Frontier lab AI solutions are costly, resource-intensive, and ethically/legally under scrutiny. The challenge lies in demonstrating how Apertus can help to the gap between digitized documents and AI-usable knowledge graphs that can reveal patterns across cases, enable semantic search, and preserve the documentary evidence for academic integrity. By developing an open, systematic pipeline, we are helping to fully prepare datasets for model training and further analysis, while preserving their historical integrity and enabling wider understanding. Our local advantage: the University of Bern has research outputs that already define a target structure for archivists, so they can serve as a reference against which automated pipelines are evaluated.
Main Objective
Develop an end-to-end, reproducible data pipeline — powered by Apertus 1.5 — that processes historical Swiss archival documents into AI-ready data assets while maintaining document origin trails. Your solution should combine automatic document structuring (titling, pagination hygiene, flagging of handwritten/illegible pages), metadata enrichment, and bias-aware preprocessing to create a dataset suitable for training Apertus (or similar language models) on Swiss legal and historical topics. Evaluate your pipeline against a University of Bern case study, and demonstrate value with a small downstream application — e.g., dossier Q&A with page citations, procedural timeline extraction, or an MCP server. See the slides in the fileshare for a visual example.
2. 2
Datasets (type, format, access)
Use one or more datasets available through opendata.swiss or the University of Bern research repositories. Three examples are linked, by increasing reach: 1. **Stapfer-Enquête**: A unique survey of the school situation in the Helvetic Republic, initiated in 1799 by education minister Philipp Albert Stapfer (1766–1840). The ~2,400 handwritten response sheets are held in the Swiss Federal Archives (Archiv der Helvetik, Bestand B0#C.02.3.1). A University of Bern team datafied the survey 2009–2015 and handed it to the Archives, which operate www.stapferenquete.ch - Type/format: handwritten page scans plus the published datafication (transcriptions/structured records). Access: opendata.swiss 2. **Federal Consultations 1960–1991**: Metadata of the Federal Chancellery’s online index of consultations - Type/format: structured tabular metadata. Access: opendata.swiss 3. **Archive Database of the Swiss Federal Archives**: Metadata of originally 3.84 million dossiers and 257,000 documents, updated in February 2026. Type/format: structured metadata / linked data (RDF) / images. Access: opendata.swiss All three datasets are published on opendata.swiss — check and respect the licence stated for each dataset, and download from there. Types of preprocessing that may be needed: transcription normalization and structure extraction; schema harmonization; linked-data handling, entity linking; date/geolocation parsing; anonymization verification.
Supporting Materials
- Hack Apertus challenge fileshare - References for Hack Apertus - Focus Group AI - University of Bern - BORIS Portal - University of Bern - Apertus model technical documentation - Opendata.ch: GLAMhack repository - Swiss open data for AI assistants **Inspiring Papers & Tools**: https://www.archipanion.com/en/ https://allenai.org/blog/asta-datavoyager https://www.archivelm.com/ https://about.jstor.org/products/digital-stewardship/seeklight/ Wu J., Jiang R. & Li Z. Reinforcement learning-guided archival question answering with adaptive curriculum learning. Sci Rep (2026). https://doi.org/10.1038/s41598-026-67621-8 Nuno Miguel, Teixeira Sousa. The algorithm of silence: artificial intelligence, archival bias and the ethical reconstruction of digital memory. Journal of Documentation (2026) 82 (2): 233–248. https://doi.org/10.1108/JD-08-2025-0249 AI assistants in the archive and the lure of “instant history”. Finola Finn, Donal Khosrowi. Cambridge Forum on AI: Culture and Society (2026). https://doi.org/10.1017/cfc.2025.10012 Inventory of NARA Artificial Intelligence (AI) Use Cases https://www.archives.gov/ai The Swiss Federal Archives > Strategy 2026–2030 https://www.bar.admin.ch/en/strategy
3. 3
Preferred Programming Languages / Frameworks
Python (Pandas, Transformers) preferred for data processing and model interaction, with Gradio/Jupyter/JavaScript for dashboard interfaces. Any cloud platforms (Swiss based / Educational access) and local development environments can be used, as long as they are freely accessible to all team members.
Recommended Platforms
The Opendata.swiss portal showcase includes several exemplary technical implementations, such as Swiss open data for AI assistants – an independent open-source project (GitHub, MIT licence) by Hayal Oezkan that enables AI assistants to query official Swiss federal and cantonal data sources exposed through 42 lightweight Model Context Protocol (MCP) servers acting as read-only clients.
Specific AI Models
Apertus v1.5 — 8B or 70B
Model Sources
Official Swiss AI Initiative model weights or verified provider API, with required licenses/compliance documentation. Participants may experiment with other LLM variants if relevant for documentation contexts and baseline comparison, and will need to cite all model sources appropriately.
4. 4
Expected Outputs
- Complete pipeline code repository (GitHub/GitLab) with clear installation instructions - Generated AI-ready dataset in specified format (CSV/JSON with metadata) - Demonstration notebook showing pipeline execution with sample documents - Documentation including API documentation, usage examples, and technical rationale - Ethical impact assessment document covering data privacy and bias considerations
Success Criteria
- **Data Usability**: Pipeline correctly processes 80%+ of documents with <5% structural errors - **AI Friendliness**: Output includes model-ready JSON/TSV with standardized metadata schema - **Provenance Preservation**: Clear documentation of editing decisions and original document references - **Performance**: Pipeline completes within 2 hours for 1000+ page dataset on standard machine spec - **Interoperability**: Output conforms to shared specification ensuring compatibility across submissions
Evaluation Criteria
- **Technical depth**: Complexity and robustness of data pipeline - **AI readiness**: Quality of output format and metadata for model consumption - **Ethical compliance**: Documentation of privacy preservation, anonymization methods, and bias mitigation - **Innovation**: Novel approaches to handling archival document challenges - **Documentation & presentation**: Clarity and completeness of submission materials
5. 5
Ethical Guidelines
- Adherence to Swiss data protection laws and archival restrictions. - Original document versions should be preserved alongside processed versions if possible. Clear documentation of all data modifications with annotations explaining why changes were made. - The Apertus Charter (https://apertus-ai.org/pages/charter/) may be used as a reference to include assessment of potential biases in data selection, preprocessing, and model training recommendations.
Legal & Compliance Requirements
- Respect all copyright limitations established by Swiss Federal Archives for public documents vs. restricted access materials. Consider fair use / research exceptions in documentation. - Follow all conditions governing access to archival materials (some documents may have restricted access periods), use public bulk open data, avoid crawling and scraping. - All work must cite original documents and should contribute positively to open scholarship. Any trained models based on this data should consider appropriate licensing (Apache 2.0 recommended by Apertus)