Datasets:
File size: 5,627 Bytes
309f2c6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 | ---
license: cc-by-4.0
language:
- de
task_categories:
- text-generation
- fill-mask
- summarization
size_categories:
- 10K<n<100K
pretty_name: German Municipal Council Documents (Ratsvorlagen)
tags:
- german
- long-context
- public-administration
- government
- oparl
- open-government-data
- administrative-language
configs:
- config_name: default
data_files:
- split: train
path: ratsvorlagen_*.jsonl.zst
---
# German Municipal Council Documents (Ratsvorlagen-Korpus)
Full texts of municipal council documents from four German cities, prepared for
training and evaluating language models. **43,127 documents, 302.9 million tokens**
(cl100k_base).
Mirror of the archived release on Zenodo:
[10.5281/zenodo.21767760](https://doi.org/10.5281/zenodo.21767760). The Zenodo record
is the citable version of record; this repository exists so the data is usable where
people actually look for German training corpora.
## Why this corpus exists
German long-context material with a permissive licence is scarce. Municipal council
documents are coherent, domain-dense administrative prose and, unlike most
parliamentary corpora, largely unexplored. Included are council proposals
(Beschlussvorlagen), motions (Anträge), formal enquiries (Anfragen) and information
items (Informationsvorlagen), always the public part only.
## What you get, and what you do not
Read this before you plan an experiment around it:
- **Leipzig dominates.** 62 % of all documents come from one city.
- **Six budget plans carry about 10 % of all tokens.** A handful of files skews any
token-weighted statistic.
- **The longest documents are numeric tables, not prose.** If you need long-form
German prose, the length distribution will mislead you. The single longest document
has 7,375,767 tokens across 10,189 pages and is a budget appendix.
- 303 million tokens is small for pretraining. This is domain data and evaluation
material, not a pretraining corpus on its own.
| City | Documents | Tokens | Longest (tokens) | Licence |
|---|---|---|---|---|
| Leipzig | 26,685 | 200.3 M | 7,375,767 | CC BY 4.0 |
| Osnabrück | 10,334 | 77.2 M | 676,664 | CC BY 4.0 |
| Dresden | 4,064 | 3.3 M | 100,808 | DL-DE/Zero 2.0 |
| Freiburg im Breisgau | 2,044 | 22.1 M | 566,107 | DL-DE/BY 2.0 |
Length distribution: 5,082 documents above 8K tokens, **1,540 above 32K**, **388 above
100K**. Duplicates across the whole corpus: 35 exact (SHA-256 over normalised text) and
265 near-duplicates (MinHash, 128 permutations, Jaccard >= 0.8).
## Fields
JSONL, one document per line, UTF-8, zstd-compressed per city.
| Field | Meaning |
|---|---|
| `city_slug` | city identifier |
| `oparl_id` | identifier in the source system |
| `name`, `reference` | title and official reference number |
| `date`, `paper_type` | date and document type |
| `source_url` | the exact source document |
| `pdf_bytes`, `pages` | size and page count of the source PDF |
| `text` | extracted full text |
| `tokens_cl100k` | token count (cl100k_base) |
| `masked` | whether contact data was replaced in this document |
| `rolle`, `datei_name` | Freiburg and Dresden only: main document or attachment |
`INDEX.csv.zst` holds all 43,127 documents without the text, so distributions can be
recomputed without downloading the archives. `PREVIEW.jsonl` contains 11 complete
documents spanning the length classes. `STATS.json` has every figure in machine-readable
form.
## Provenance
Collected through the official OParl endpoints of the cities, source snapshot of
2026-07-29, retrieved 2026-07-31, 2026-08-01 and 2026-08-02, at one request per second
per host (Dresden one per three seconds) with an identifying user agent. Retrieval
success rates: Leipzig 85.7 %, Osnabrück 99.8 %, Dresden 41.9 %, Freiburg im Breisgau
94.4 %. No access restriction was circumvented; all endpoints are anonymously
accessible.
## Privacy
Contact details are masked: e-mail addresses, telephone and fax numbers and IBANs were
replaced with placeholders (3,799 of 43,127 documents were affected). **Personal names
are kept**, because council members and municipal staff act in an official capacity
here and the texts lose their meaning without them. This is a deliberate decision, not
an oversight.
## Licence and attribution
The collection is CC BY 4.0, the strictest of the source licences. Each document
carries its city's licence, verified directly at the OParl system endpoint. Please
credit Stadt Leipzig, Stadt Osnabrück, Landeshauptstadt Dresden and Stadt Freiburg im
Breisgau as sources of the respective documents.
## Citation
```bibtex
@dataset{cherif_2026_ratsvorlagen,
author = {Cherif, Khaled},
title = {Ratsvorlagen-Korpus deutscher Kommunen: 43.127 Dokumente
Long-Context-Verwaltungstext},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.21767760}
}
```
---
## Deutsche Kurzfassung
Volltexte kommunaler Ratsdokumente aus Leipzig, Osnabrück, Dresden und Freiburg im
Breisgau: 43.127 Dokumente, 302,9 Mio. Tokens, CC BY 4.0, über die OParl-Schnittstellen
der Städte erhoben. 1.540 Dokumente liegen über 32K Tokens, 388 über 100K.
Ehrlich zu den Grenzen: Leipzig stellt 62 Prozent der Dokumente, sechs Haushaltspläne
tragen rund 10 Prozent aller Tokens, und die längsten Dokumente sind Zahlentabellen,
keine Prosa. Kontaktdaten sind maskiert, Personennamen in amtlicher Funktion bleiben
bewusst erhalten.
Zitierfassung mit DOI: https://doi.org/10.5281/zenodo.21767760
Entstanden im Rahmen von [InfraNode](https://infranode.dev), einer schlüssellosen
offenen Datenschnittstelle für deutsche Städte.
|