Diavgeia — Greek Government Transparency Decisions

Dataset Info

This dataset contains the full text and metadata of public-sector decisions (αποφάσεις / πράξεις) published on Diavgeia (diavgeia.gov.gr), the Greek government's transparency portal. Since 2010 (Law 3861/2010), every Greek public entity is legally required to publish its administrative acts — budget commitments, expenditure approvals, contracts, appointments, regulatory acts, and more — on Diavgeia, each identified by a unique ADA (Αριθμός Διαδικτυακής Ανάρτησης).

Each record corresponds to a single decision. The original PDF document was collected together with its structured metadata, and the document body was extracted and converted to clean Markdown text following the glossAPI processing pipeline.

This dataset was produced as part of the glossAPI project of EELLAK (Open Technologies Alliance), which builds high-quality, openly licensed Greek-language text corpora for AI/NLP research, linguistic analysis, and open data reuse.

Source and Collection

• Source: Diavgeia OpenData API — https://diavgeia.gov.gr/opendata/search
• Document endpoint: https://diavgeia.gov.gr/doc/{ADA}
• Collection period: January–February 2026
• Identifier: each decision is keyed by its ADA
• Original format: PDF + JSON metadata, OCR/text-extracted to Markdown

The raw harvest comprised roughly 3.1 million decision documents (~1.26 TB of PDFs). This release publishes the 1,840,305 records that were successfully extracted into clean Markdown text and passed quality filtering.

Important note on temporal coverage

The Diavgeia OpenData search API silently applies a default ~6-month rolling window on both submissionTimestamp and issueDate when no explicit date parameters are supplied. As a result, this collection is concentrated in the ~6-month window preceding the scrape (approximately mid-2025 to early 2026), rather than the complete 2010–present archive (estimated at ~70M+ decisions). Older decisions remain reachable through the API only by walking the from_issue_date parameter in ~6-month increments. Users should treat this dataset as a large recent snapshot, not a full historical census. See Limitations below.

Dataset Structure

The dataset is provided as Parquet with a single train split.

Column	Type	Description
id	string	The decision's unique ADA (Αριθμός Διαδικτυακής Ανάρτησης)
markdown_text	string	The extracted document body, converted to clean Markdown text
metadata_json	string	A JSON string with the decision's structured metadata (schema below)

metadata_json schema

The metadata_json field is a JSON-encoded object containing the official Diavgeia metadata for the decision:

Field	Description
subject	Title / subject of the decision (θέμα)
protocolNumber	Internal protocol number of the issuing body
issueDate	Issue date, Unix epoch in milliseconds
organizationId	ID of the issuing organization / public entity
signerIds	List of IDs of the signer(s)
unitIds	List of organizational unit IDs
decisionTypeId	Decision type code (see Decision Types below)
thematicCategoryIds	List of thematic category IDs
extraFieldValues	Type-specific structured fields (e.g. financial year, amounts, CPV, KAE)

Note: dates are stored as milliseconds since the Unix epoch. To convert in Python: datetime.utcfromtimestamp(issueDate / 1000).

Example usage

import json
from datasets import load_dataset

ds = load_dataset("glossAPI/diavgeia", split="train")

row = ds[0]
print(row["id"])                 # e.g. "ΨΓΠ9469ΗΡ8-ΧΚ5"
print(row["markdown_text"][:500])

meta = json.loads(row["metadata_json"])
print(meta["subject"], meta["decisionTypeId"], meta["issueDate"])

Decision Types (Categorizations)

Diavgeia organizes decisions into a hierarchical taxonomy. Our research on the API identified 43 decision types in total — 35 with documents and 8 empty parent categories (organizational nodes with no documents of their own: 2.4.1, 2.4.2, 2.4.3, 2.4.4, 2.4.5, 2.4.6, 2.4.7, 2.4.8).

34 decision types are represented in this dataset. The distribution is highly skewed — the top 3 types account for roughly two-thirds of all records. The table below shows the main types observed, with their share of the collected corpus:

Type code	Description (Greek)	Approx. share
Β.2.2	Ανάληψη Υποχρέωσης	~33.8%
Β.1.3	Πράξη μεταβολής Ανθρώπινου Δυναμικού	~21.8%
2.4.7.1	Λοιπές ατομικές διοικητικές πράξεις	~12.3%
Β.2.1	Ανάθεση / Κατακύρωση	~7.2%
Α.2	Κανονιστική Πράξη	~5.6%
Δ.1	Σύμβαση	~5.0%
Γ.3.4	Πράξη ΟΤΑ (λοιπές ατομικές)	~4.6%
2.4.6.1	Αποφάσεις / Πράξεις Συλλογικών Οργάνων	~2.6%
Β.1.1	Πράξη Διορισμού / Πρόσληψης	~2.6%
Γ.2	Πράξη ΟΤΑ (οικονομικές)	~1.1%
Δ.2.2	Εντολή Πληρωμής / Κατακύρωση	~0.8%
Β.1.2	Λύση Υπαλληλικής Σχέσης	~0.7%
Δ.2.1	Δαπάνη / Περίληψη Διακήρυξης	~0.5%
Ε.4	Περίληψη Διακήρυξης / Αναπτυξιακού Νόμου	~0.2%
Γ.3.1	Προκήρυξη Πλήρωσης Θέσεων	~0.2%
Γ.3.2	Πίνακες Επιτυχόντων / Κατάταξης	<0.2%
—	plus ~18 additional long-tail types	~0.9%

Full taxonomy of decision types found in the API (35 with documents)

Β.2.2, Β.1.3, 2.4.7.1, Β.2.1, Α.2, Δ.1, Γ.3.4, 2.4.6.1, Β.1.1, Γ.2, Δ.2.2, Β.1.2, Δ.2.1, Ε.4, Γ.3.1, Γ.3.2, Β.3, Β.5, Β.4, Γ.3.3, Γ.3.5, 100, Α.3, Ζ.1, Ε.1, Ε.2, Ε.3, Α.4, Α.5, Α.7, Α.1.2, Γ.3.6, Α.6, Α.1.1, Β.6

Dataset Statistics

Counts cover the markdown_text field, following the glossAPI convention.

• Records (rows): 1,840,305
• Size on disk: ~3.2 GB (Parquet)
• Decision types represented: 34
• Document length (markdown_text): ranges from ~150 to ~6 million characters
• Words: to be computed (whitespace-separated, on markdown_text)
• Tokens: to be computed (nlpaueb/bert-base-greek-uncased-v1 WordPiece tokenizer, no special tokens, no truncation)

Intended Uses

• Pre-training and fine-tuning of Greek-language models
• Domain adaptation for legal / administrative / public-sector Greek
• Text classification (e.g. by decision type or thematic category)
• Summarization and information extraction from administrative documents
• Linguistic and computational-social-science research on Greek public administration

Limitations and Biases

• Temporal coverage: concentrated in a ~6-month recent window (see note above); not a complete historical archive of Diavgeia.
• Type skew: financial/administrative types (Β.2.2, Β.1.3, 2.4.7.1) dominate, and many of these are highly templated/boilerplate documents. This may bias models toward repetitive administrative language.
• OCR / extraction noise: text is derived from PDFs and may contain OCR artifacts, broken character encodings, table-flattening, or layout loss — particularly in scanned or image-based documents.
• Personal data: decisions are official public-transparency records and may contain personal data (names, IDs) that were lawfully published by the issuing bodies. Use responsibly and in compliance with applicable law (incl. GDPR).
• Coverage gaps: ~3.1M documents were harvested; this release contains the subset that was successfully converted to clean text.

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY-4.0) license. The underlying decisions are official public documents published on Diavgeia under the Greek transparency framework.

Citation

If you use this dataset, please cite:

@misc{glossapi_diavgeia,
  title        = {Diavgeia — Greek Government Transparency Decisions},
  author       = {{glossAPI Team, EELLAK (Open Technologies Alliance)}},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/glossAPI/diavgeia}
}

Acknowledgements

• Diavgeia — Greek transparency portal: https://diavgeia.gov.gr
• OpenData API: https://diavgeia.gov.gr/api/help

Contact

For feedback contact: glossapi.team@eellak.gr