AndreaBozzo/ceres-open-data-index

Ceres Open Data Index

A curated index of 1,854,596 open datasets from 47 open data portals across ~25 countries plus international sources (EU Open Data Portal, HDX), with cross-portal duplicates flagged. This is the largest aggregated open data metadata index available as a single downloadable resource.

Dataset Description

This dataset contains normalized metadata from government and institutional open data portals. It was harvested by Ceres, a harvest-first toolkit for open data, via CKAN, DCAT-AP (udata REST), and SPARQL endpoints. Metadata has been flattened into a tabular schema, with noise filtered and cross-portal duplicates flagged (flagged, not removed).

What's included

• Dataset titles, descriptions, URLs, tags, organizations, and licenses from 47 open data portals
• A heuristic cross-portal duplicate signal (is_duplicate flag) — a title match, not canonical deduplication (see Curation)
• Portal-level language annotations
• Per-portal Parquet subsets for targeted analysis
• A versioned snapshot manifest, coverage/quality report, and snapshot-to-snapshot changelog

What's NOT included

• The actual data files (CSVs, shapefiles, etc.) -- only metadata
• Embedding vectors (model-specific; regenerate with any embedding model)
• Resource-level metadata (individual file URLs within datasets)

Note: SPARQL-sourced rows (data.europa.eu) carry title, description, identifier, and modification date; portal-specific fields such as tags/organization/license are populated mainly for CKAN/udata sources.

Dataset Structure

Schema

Column	Type	Nullable	Description
original_id	string	no	Dataset ID from the source portal
source_portal	string	no	Portal base URL
portal_name	string	no	Human-readable portal name
url	string	no	Direct URL to the dataset page
title	string	no	Dataset title
description	string	yes	Dataset description
tags	string	yes	Comma-separated tag names
organization	string	yes	Publishing organization
license	string	yes	License title or identifier
metadata_created	string	yes	Original creation date (ISO 8601)
metadata_modified	string	yes	Last modification date (ISO 8601)
first_seen_at	string	no	When Ceres first indexed this dataset (RFC 3339)
language	string	yes	Primary language code (e.g. it, en, de)
is_duplicate	boolean	no	Heuristic signal: the same title (case-insensitive) appears on another portal. Not canonical deduplication

Files

• all.parquet -- Complete dataset (1,854,596 rows)
• data/<portal>.parquet -- Per-portal subsets (47 files; rows repeat those in all.parquet)
• identity.parquet -- Slim per-record fingerprint (source_portal, original_id, content_hash), one row per all.parquet row, used to diff snapshots
• metadata.json -- Versioned snapshot manifest: stable snapshot_id, generation time, curation counts, per-file SHA-256 checksums, per-portal inclusion, and duplicate_detection provenance
• reports.json / report.md -- Coverage and quality report (coverage by portal/type/profile/language; field-completeness rates)
• changelog.json / changelog.md -- What changed versus the previous snapshot (added/changed/removed/unchanged, per portal)

Splits by portal

Portal	Country	Datasets
data.europa.eu	EU (international)	689,589
catalog.data.gov	USA	399,597
govdata.de	Germany	146,648
data.gov.au	Australia	135,267
dati.gov.it	Italy	64,483
ckan.publishing.service.gov.uk	United Kingdom	56,433
open.canada.ca	Canada	47,282
datos.gob.es	Spain	46,998
data.gov.ua	Ukraine	36,866
data.humdata.org	International	27,795
data.gouv.fr	France	21,059
data.gov.ie	Ireland	20,998
data.overheid.nl	Netherlands	20,679
dados.gov.pt	Portugal	20,528
ckan.opendata.swiss	Switzerland	15,441
dati.toscana.it	Italy	12,172
ckan.open.nrw.de	Germany	10,281
podatki.gov.si	Slovenia	8,935
data.norge.no	Norway	8,604
catalog.data.metro.tokyo.lg.jp	Japan	8,466
data.gov.cz	Czechia	5,815
datosabiertos.gob.pa	Panama	5,577
discover.data.vic.gov.au	Australia	5,575
dati.regione.marche.it	Italy	5,448
data.gov.ro	Romania	5,003
catalogue.data.gov.bc.ca	Canada	3,343
datos.gob.cl	Chile	3,019
data.public.lu	Luxembourg	2,676
dati.comune.milano.it	Italy	2,596
dati.emilia-romagna.it	Italy	2,577
www.avoindata.fi	Finland	2,524
avoindata.suomi.fi	Finland	2,524
opendata.aragon.es	Spain	2,429
dati.puglia.it	Italy	1,814
dati.trentino.it	Italy	1,373
datos.gob.ar	Argentina	1,203
ckan0.cf.opendata.inter.prod-toronto.ca	Canada (Toronto)	548
dati.regione.umbria.it	Italy	446
dati.lazio.it	Italy	402
dati.comune.roma.it	Italy	365
dati.regione.campania.it	Italy	360
opendata-hro.de	Germany	280
dati.regione.sicilia.it	Italy	192
dati.comune.genova.it	Italy	163
dati.regione.liguria.it	Italy	123
lindas.admin.ch	Switzerland	64
dati.comune.napoli.it	Italy	36

Per-portal file names are derived from the portal hostname (e.g. data/catalog-data-gov.parquet, data/data-europa-eu.parquet, data/data-gov-au.parquet).

Coverage & quality

Figures below are generated with the snapshot (reports.json / report.md).

By source type: CKAN 1,112,076 · DCAT-AP 742,520 (of which SPARQL 698,257, udata REST 44,263).

By language (primary code per portal):

Language	Datasets		Language	Datasets
English (en)	1,395,362		Dutch (nl)	20,679
German (de)	172,714		Portuguese (pt)	20,528
Italian (it)	92,550		Norwegian (no)	8,604
Spanish (es)	59,226		Japanese (ja)	8,466
Ukrainian (uk)	36,866		Czech (cs)	5,815
French (fr)	23,735		Finnish (fi)	5,048
			Romanian (ro)	5,003

Field completeness (share of rows with a non-empty value): description 100.0% · organization 56.3% · modification_date 57.1% · tags 49.7% · license 32.7%.

Loading the dataset

import pandas as pd

# Load all datasets
df = pd.read_parquet("all.parquet")

# Load a specific portal (US federal)
df_us = pd.read_parquet("data/catalog-data-gov.parquet")

# Drop rows flagged by the heuristic duplicate signal (title match, not canonical)
unique = df[~df["is_duplicate"]]

Or with the datasets library:

from datasets import load_dataset

ds = load_dataset("AndreaBozzo/ceres-open-data-index")

Dataset Creation

Harvesting methodology

Ceres harvests each portal through the API best suited to it, with a streaming, resilient pipeline:

1. CKAN portals are paged via package_search (streaming, content-hash delta detection). The client adapts on slow/flaky portals: per-request timeout grows and page size shrinks on timeouts/truncated bodies, then recovers, so large catalogs page fully without manual tuning.
2. US data.gov is harvested via its relocated action API on api.gsa.gov (API-key authenticated); datasets stay attributed to catalog.data.gov.
3. DCAT-AP (udata REST) portals (e.g. France, Portugal, Luxembourg) are paged via the catalog JSON-LD endpoint, with rate-limit cooldown/retry.
4. data.europa.eu is currently a partial SPARQL harvest: rows with English or untagged titles are represented, while datasets that only expose titles in other languages remain out of scope. A complete, restartable harvest requires a stable upstream bulk or pagination contract.
5. Export: ceres export --format parquet streams all datasets from PostgreSQL, applies curation filters, flattens metadata, detects cross-portal duplicates, and writes compressed Parquet.

Curation

The following curation steps are applied during export:

• Noise filtering (26,640 rows removed): titles shorter than 5 characters, titles containing "test"/"prova"/"esempio", datasets with empty descriptions
• Duplicate flagging (769,517 rows marked): datasets whose title (case-insensitive) appears on more than one portal are flagged with is_duplicate=true but kept in the dataset. This is a heuristic signal, not canonical deduplication — it performs no entity resolution. The matching rule and its version are recorded in metadata.json under duplicate_detection (title-exact-ci-cross-portal, v1). Known mirrored/alternate endpoints can be declared as portal aliases so they are not counted as independent sources. Treat the flag as "likely duplicate by title", not a guarantee.
• Metadata flattening: tags, organization, and license are extracted from nested CKAN JSON into top-level string columns

Source data

All data comes from publicly accessible CKAN, DCAT-AP, and SPARQL endpoints of government-operated or institutional open data catalogs. Only the US data.gov source requires an API key (free).

Update frequency

This snapshot (ceres-20260626-87b9b74e080d) was generated on 2026-06-26. The Ceres harvester can refresh the index at any time; each refresh publishes a new manifest, report, and changelog.

Considerations

Known biases

• Geographic skew: the EU Open Data Portal (data.europa.eu) is the largest single source (~37%), followed by the USA (catalog.data.gov, ~22%), Germany (govdata.de + NRW), and Australia (federal + Victoria). Italy contributes many small regional/municipal portals.
• data.europa.eu coverage: the snapshot contains 689,589 rows from a portal reporting roughly 1.98M datasets. Datasets titled only in languages other than English are under-represented.
• Language distribution: predominantly English (~75%), followed by German, Italian, Spanish, Ukrainian, French, Dutch, Portuguese, Norwegian, Japanese, Czech, Finnish, and Romanian. language is the portal's primary language, not per-row language detection.
• Aggregation duplicates: national aggregators (e.g. dati.gov.it, data.europa.eu) re-list datasets from regional/member portals, so a large share of rows are flagged as cross-portal duplicates.

Limitations

• Metadata quality varies by portal -- some datasets have minimal descriptions or missing fields
• Tags and organization names are not standardized across portals
• metadata_created/metadata_modified are portal-reported and may not reflect actual data updates
• License information is inconsistent across portals
• SPARQL-sourced rows (data.europa.eu) carry fewer columns (title/description/dates) than CKAN/udata rows

Ethical considerations

This dataset contains only public metadata (titles, descriptions, tags) from government open data portals. It does not contain personal data. All source portals publish data under open licenses (typically CC-BY, CC0, ODbL, or equivalent).

Additional Information

License

This dataset is released under Apache 2.0, consistent with the Ceres project. The underlying metadata is sourced from portals that publish under open licenses.

Citation

@misc{ceres-open-data-index-2026,
  title={Ceres Open Data Index},
  author={Andrea Bozzo},
  year={2026},
  url={https://github.com/AndreaBozzo/Ceres},
  note={Aggregated metadata from 47 open data portals, snapshot 2026-06-26}
}

Links

• Ceres on GitHub
• CKAN API documentation