Riksrevisjonen/sai-mash

Multilingual Audits: Structured & Harmonized (MASH)

MASH is a dataset of Supreme Audit Institution (SAI) reports harmonized to a common language and format. SAIs publish their work in national languages and with varying structures, making cross-country analysis difficult. MASH resolves this by processing each report through a standardized pipeline that produces English summaries, structured metadata, and controlled-vocabulary tags — enabling researchers, auditors, and developers to treat reports from different countries as a unified, queryable corpus.

---

Use Cases

• QA systems: Ask questions across the full corpus: "Has any SAI audited public procurement in the last five years? What methodologies did they use?"
• Cross-country comparison: Identify which topics receive audit attention in which countries, and how audit approaches differ.
• Research and benchmarking: Study audit coverage, methodology trends, and thematic gaps across institutions.

---

Coverage

Country	Institution
Brazil	Tribunal de Contas da União (TCU)
Finland	Valtiontalouden tarkastusvirasto (VTV)
Netherlands	Algemene Rekenkamer
Norway	Riksrevisjonen
United Kingdom	National Audit Office (NAO)

---

How the Dataset Was Created

Reports are processed through a five-stage pipeline:

1. OCR — Raw PDFs are converted to Markdown using a document conversion API, preserving text structure.
2. Summarize & translate — Each Markdown file is sent to an LLM that produces a structured English summary with sections for background, methodology, findings, recommendations, and conclusion. This normalizes both language and format regardless of the original.
3. Topic and methodology extraction — The LLM reads each summary and assigns up to 10 topics (policy area or sector audited) and up to 10 methodologies (audit approaches used). Both lists draw from a shared controlled vocabulary maintained across all countries, ensuring consistent terminology for cross-country queries.
4. Year detection — The release year is extracted from the filename when unambiguous; otherwise the LLM reads the document head and tail to identify the publication date.
5. Collection — All outputs are assembled into a single Parquet file. Only fully processed rows — those with a summary, topics, methodologies, and year — are included by default.

---

Schema

Column	Type	Description
filename	string	Original Markdown filename, encoding country, a document ID, and the publication date.
markdown	string	Full OCR'd content of the report in its original language.
title	string	English title of the report, extracted from the first line of the summary.
summary	string	Structured English summary produced by the LLM, covering background, methodology, findings, recommendations, and conclusion.
topics	string	Comma-separated list of subject topics drawn from a shared controlled vocabulary (e.g. Tax administration, Public procurement).
methodologies	string	Comma-separated list of audit methodologies drawn from a shared controlled vocabulary (e.g. Compliance audit, Data analytics, Document review and interviews).
year	int32	Report release year. Nullable — absent if the year could not be determined with confidence.