SinclairSchneider/zeit_de

DIE ZEIT / ZEIT ONLINE (zeit.de)

342 722 German-language news articles crawled from www.zeit.de and published as a Hugging Face dataset. This is part of the German Media collection, a set of comparable corpora covering the German media landscape from public-service broadcasters to alternative and fringe outlets, all built with the same crawler and the same extraction pipeline so that they can be compared and combined directly.

⚠️ For scientific research only

This corpus is published exclusively for scientific research under the German research exception for text and data mining — the Forschungsprivileg of § 60d UrhG (implementing Art. 3 of EU Directive 2019/790). The article texts remain the copyright of their publisher and are not licensed for commercial use, redistribution, or republication. By downloading it you confirm that you use it for non-commercial scientific research. See Licensing and intended use.

The articles are reproduced verbatim from the publisher's website; their content expresses the views of that publisher, not of the dataset author.

At a glance

Source	www.zeit.de
Rows	342 722
Columns	10 (title, content, author, description, keywords, category, news_type, datePublished, dateModified, url)
Language	German (de)
Split	train (single split)
Format	Parquet, 3 shards, ≈ 732 MB download / 1,240 MB in memory
Perceived partisanship	3.6 / 7 → relatively neutral (medienkompass.org reader survey)
Intended use	Scientific research only (Forschungsprivileg, § 60d UrhG)
Mirror	NLP-UniBW/zeit_de

from datasets import load_dataset

ds = load_dataset("SinclairSchneider/zeit_de", split="train")
print(ds[0]["title"])

About the source

DIE ZEIT is a German national weekly newspaper founded in 1946 in Hamburg. It is published by the Zeitverlag Gerd Bucerius, part of Dieter von Holtzbrinck Medien, and is known for long-form reporting, essays and debate; editorially it is usually described as liberal / left-liberal. zeit.de (ZEIT ONLINE) has its own newsroom and publishes daily news alongside the weekly paper's articles; content marked Z+ is reserved for subscribers and is not contained in this dataset.

Perceived political position

The German media compass at medienkompass.org places German mainstream and alternative outlets on two axes on the basis of a reader survey (n ≈ 1 154 participants in total, building on a media map by the think tank polisphere; the quality axis is adapted from Vanessa Otero's media chart).

• x axis — Parteilichkeit (partisanship), 1–7: < 3.5 left, 3.5 – 4.5 relatively neutral, > 4.5 right. 1 far-left (fake news and conspiracy theories) · 2 left mission (questionable journalistic standards) · 3 leaning left · 4 minimal partisan tendency · 5 leaning right · 6 right mission (questionable journalistic standards) · 7 far-right (fake news and conspiracy theories).
• y axis — Qualität (quality), 1–5: 1 sensationalist / clickbait · 2 basic information · 3 meets high standards · 4 analytical · 5 complex.

In that survey, Zeit is rated:

	Score	Reading
Partisanship (1–7)	3.6	relatively neutral
Quality (1–5)	3.5	on the 1 = clickbait … 5 = complex scale

Parteilichkeit (partisanship):  1 = far left  ...  7 = far right

1         2         3         4         5         6         7
|---------|---------|---------|---------|---------|---------|
                          ^                                  
                        Zeit = 3.6

left  <--------------|  neutral  |-------------->  right
                   3.5           4.5

For context, here is the full survey table, sorted from left to right — the other outlets in the collection are directly comparable:

Outlet	Partisanship (1–7)	Quality (1–5)
indymedia	2.0	1.7
Jungle World	2.3	2.3
junge Welt	2.4	2.3
neues Deutschland	2.6	2.3
der Freitag	2.7	2.9
die Tageszeitung	2.8	2.7
Vice	2.8	1.8
NachDenkSeiten	3.1	3.0
Frankfurter Rundschau	3.4	2.8
SPON	3.5	2.1
Spiegel Print	3.5	2.9
Süddeutsche	3.5	3.3
WDR	3.5	2.8
Tagesspiegel	3.6	2.8
Zeit ⬅ this dataset	3.6	3.5
ARD	3.7	2.9
NDR	3.7	2.8
Deutschlandfunk	3.8	3.5
Stern	3.8	2.1
ZDF	3.8	2.8
MDR	4.1	2.5
Handelsblatt	4.3	3.3
NTV	4.3	2.1
BR	4.4	2.6
N24	4.4	2.0
Wirtschaftswoche	4.4	3.1
Frankfurter Allgemeine	4.5	3.3
RTL	4.5	1.5
Die Welt	4.8	2.5
Cicero	4.9	3.1
Focus Online	4.9	1.7
Deutsche Wirtschafts Nachrichten	5.0	2.2
MM News	5.1	1.7
RT Deutsch	5.1	2.0
Achgut	5.2	2.4
BILD	5.2	1.3
Tichys	5.5	2.6
junge Freiheit	5.8	1.9
Compact	6.0	1.8

How to read these numbers. They are perception data from a self-selected online survey, not a content analysis: participants rated outlets they know, the sample is not representative of the German population, and the site itself describes the project as an early iteration. Treat them as a rough, crowd-sourced orientation — useful for grouping datasets into coarse political blocks, not as ground truth for bias labelling.

Dataset structure

Single train split with 342 722 rows and the following columns:

Column	Type	Description	Extracted from
title	string	Headline of the article.	headline from ld+json, HTML-unescaped
content	string	Full plain-text body of the article.	articleBody from ld+json, HTML-unescaped
author	string	Author(s) as published by the outlet.	author.name from ld+json — only the first author when several are listed
description	string	Teaser / meta description of the article.	description from ld+json, HTML-unescaped
keywords	list[string]	Topic keywords / tags assigned by the publisher.	keywords from ld+json, split on , and stripped
category	string	Section the article was published in.	the first path segment of mainEntityOfPage.@id with every non-letter removed (e.g. politik, zeitwissen)
news_type	string	schema.org @type of the page (e.g. Article).	@type from ld+json (e.g. Article)
datePublished	string	Publication timestamp.	datePublished from ld+json
dateModified	string	Timestamp of the last editorial update.	dateModified from ld+json
url	string	Canonical URL the article was fetched from.	the crawled URL

Example record

{
  "title": "Empfehlung der Ständigen Impfkommission: Heißt das jetzt impfen oder nicht?",
  "content": " Viele Eltern und Ärztinnen mögen sich ein klares Votum der Ständigen Impfkommission (Stiko) gewünscht haben. Das gab es aber nicht. Kindern mit Vorerkrankungen empfiehlt die Stiko, sich gegen das Coronavirus immunisieren zu lassen . Zudem denjenigen, die Kontakt mit Menschen haben, die schwer an Covid-19 erkranken könnten. Alle anderen können nur auf \"individuellen Wunsch\" der Eltern und nach ärz …",
  "author": "Ingo Arzt",
  "description": "Die Stiko empfiehlt kranken Kindern zwischen fünf und elf Jahren die Corona-Impfung. Der Rest bekommt sie auf individuellen Wunsch. Was Eltern jetzt wissen müssen",
  "keywords": [
    "Johanna Schoener",
    "STIKO",
    "Pfizer",
    "BioNTech",
    "EMA",
    "EU",
    "Impfung",
    "Nebenwirkung",
    "Impfstoff",
    "Corona-Impfung"
  ],
  "category": "gesundheit",
  "news_type": "Article",
  "datePublished": "2021-12-09T20:11:29+01:00",
  "dateModified": "2021-12-09T20:11:29+01:00",
  "url": "https://www.zeit.de/gesundheit/2021-12/stiko-empfehlung-corona-impfung-kinder?&icode=01w0040k0059Inhemanew2104&utm_medium=email&utm_source=elbvertiefung&utm_campaign=zplus&utm_content=&wt_zmc=emanew.int.zabo.elbvertiefung.zplus....x"
}

Temporal coverage

The datePublished column could be parsed for 342 604 of 342 722 rows (100.0 %). Those rows span 1964-02-14 → 2025-06-06.

Year	Articles
1964	2	█
1968	2	█
1970	184	█
1994	2	█
1995	9	█
1996	10	█
1997	36	█
1998	96	█
1999	45	█
2000	77	█
2001	15	█
2002	47	█
2003	842	█
2004	709	█
2005	808	█
2006	5 720	████
2007	4 955	███
2008	6 613	█████
2009	13 462	█████████
2010	12 408	█████████
2011	12 168	████████
2012	15 016	██████████
2013	17 180	████████████
2014	17 341	████████████
2015	16 214	███████████
2016	17 163	████████████
2017	16 215	███████████
2018	17 122	████████████
2019	17 510	████████████
2020	19 490	██████████████
2021	21 921	███████████████
2022	23 590	████████████████
2023	26 886	███████████████████
2024	34 595	████████████████████████
2025	24 151	█████████████████

How the data was collected

The whole pipeline is contained in this repository under scraping/ and consists of two steps: a distributed crawler that walks the site and writes one JSON file per article, and a consolidation notebook that merges those files into the dataset you are reading.

1. Crawling — scraping/crawler.py

The crawler is a breadth-first whole-site crawler built on MechanicalSoup and multiprocessing:

1. Sharded frontier. 32 worker processes each own one multiprocessing.Manager().Queue(). A URL is assigned to a worker by int(md5(url).hexdigest(), 16) % 32, so every URL is always handled by the same worker and no locking or cross-process deduplication is needed.
2. Seed. The crawl is seeded with a single article URL (https://www.zeit.de/wirtschaft/unternehmen/2025-06/steuern-entlastungspaket-unternehmen-bund-laender) and discovers everything else by following links.
3. Link discovery. Every fetched page is parsed with BeautifulSoup/lxml; all <a href> values that start with https://www.zeit.de or with / (which are then made absolute) are added to the frontier. URLs containing # are dropped. Here base_url is https://www.zeit.de without a trailing slash, so relative links are joined without a double slash.
4. URL filtering. Before fetching, a URL is skipped when it contains /files/, /video/, .jpg, .pdf or .epub.
5. Fetching. Each URL gets up to 1000 attempts with a 10 s timeout.
6. Extraction. get_content() turns the HTML into a record (see below). Only pages with a non-empty content are written out, as pages/<flattened-url>.json.
7. Resumability. Each worker appends every known and every visited URL to data/thread_<id>_pages.txt and data/thread_<id>_pages_visited.txt, and re-reads them on start, so an interrupted crawl continues where it stopped. Creating a file named killall.txt in the working directory stops all workers cleanly.

Crawler settings used for this site:

• Worker processes: 32 (iNumberOfThreads = 32)
• Retries per URL: 1000, with a 10-second socket timeout per attempt
• Proxy pool: not used — the original crawl connected directly (the published script can still route through proxies via SCRAPER_PROXIES, see below)
• User-agent rotation: no, MechanicalSoup's default UA
• Politeness delay: none — throttling was handled by the proxy pool / retry loop
• Bot-challenge handling: none
• Raw HTML dump: disabled
• Site-specific: no minimum path depth filter
• Site-specific: the whole fetch–extract–store step is wrapped in try/except; a URL that raises an error is marked as visited and skipped

Paywall handling. Pages that contain <aside id="paywall"> (Z+) are skipped, as are pages without any ld+json block.

2. Article extraction

get_content() reads the fields from the page's JSON-LD (<script type="application/ld+json">) and/or its HTML, as follows:

Field	Source in the page
title	headline from ld+json, HTML-unescaped
content	articleBody from ld+json, HTML-unescaped
author	author.name from ld+json — only the first author when several are listed
description	description from ld+json, HTML-unescaped
keywords	keywords from ld+json, split on , and stripped
category	the first path segment of mainEntityOfPage.@id with every non-letter removed (e.g. politik, zeitwissen)
news_type	@type from ld+json (e.g. Article)
datePublished	datePublished from ld+json
dateModified	dateModified from ld+json
url	the crawled URL

A ZEIT page carries several ld+json blocks. For every field the parser takes, among the blocks that contain that key, the alphabetically first one. Blocks that are not valid JSON are re-parsed with the more lenient json5 parser. When a page has no datePublished / dateModified, the placeholder 01.01.1970 is stored instead.

Each article is stored as a JSON file of the shape

{
  "title": "…",
  "content": "…",
  "additional_attributes": { "author": "…", "description": "…", "keywords": "…", "category": "…", "news_type": "…", "datePublished": "…", "dateModified": "…", "url": "…" }
}

3. Consolidation — scraping/finalize.ipynb

1. glob("pages/*.json") collects every article file.
2. The nested records are flattened into one pandas.DataFrame column per field.
3. df.drop_duplicates(subset=['title']) removes duplicates — the same article is often reachable under several URLs (print/mobile variants, tracking parameters, section mirrors), and the headline is the most reliable identity key available.
4. The frame is written to zeit_de.json and pushed with datasets.Dataset.from_pandas(df).push_to_hub("SinclairSchneider/zeit_de").

Site-specific consolidation steps for this dataset:

• Multi-page articles. ZEIT ONLINE splits long articles over several pages (…/seite-2, …) and also serves a single-page version under …/komplettansicht. Wherever a *_komplettansicht.json file exists, it is used and the first page as well as the seite-N pages of the same article are dropped. The resulting file list is written to files_combined.txt and read back from there.

Dataset-specific notes

• 01.01.1970 placeholder. When a page has no datePublished / dateModified, the crawler stores the string 01.01.1970. This is the spike of 184 rows in 1970 in the year table above (the neighbouring years 1964 and 1968 have 2 rows each); treat those dates as missing.
• author holds only the first author of co-authored articles.

Reproducing the crawl

pip install mechanicalsoup beautifulsoup4 lxml python-dateutil fake-useragent json5 \
            pandas datasets huggingface_hub tqdm

mkdir zeit_de && cd zeit_de
curl -L -O https://huggingface.co/datasets/SinclairSchneider/zeit_de/resolve/main/scraping/crawler.py

python crawler.py          # writes pages/*.json; `touch killall.txt` to stop
jupyter notebook finalize.ipynb

Routing the crawl through your own proxies

The original crawl ran behind a pool of forward proxies on a private network. Those addresses are infrastructure-specific and have been removed from the published script. Instead, crawler.py reads the SCRAPER_PROXIES environment variable: a comma-separated list of proxies, written with or without a scheme (http:// is assumed when none is given). One entry is picked per request, so a pool spreads the load across several exits.

# a single proxy
export SCRAPER_PROXIES="proxy-a.example.com:3128"

# a pool, rotated per request
export SCRAPER_PROXIES="proxy-a.example.com:3128,proxy-b.example.com:3128"

# with an explicit scheme and credentials
export SCRAPER_PROXIES="http://user:pass@proxy.example.com:3128"

# SOCKS, if PySocks is installed
export SCRAPER_PROXIES="socks5://localhost:9050"

# back to a direct connection — this is the default
unset SCRAPER_PROXIES

python crawler.py

Internally this becomes requests' standard {'http': …, 'https': …} proxy mapping; with the variable unset the mapping is empty and the crawler connects directly. No credentials, cookies or personal data are contained in the published code.

Please crawl responsibly: respect robots.txt and the site's terms of use, keep the concurrency far below the 32 workers used here unless you have the site's permission, and prefer this published dataset over re-crawling.

Usage

from datasets import load_dataset
import pandas as pd

ds = load_dataset("SinclairSchneider/zeit_de", split="train")

# stream instead of downloading everything
stream = load_dataset("SinclairSchneider/zeit_de", split="train", streaming=True)
print(next(iter(stream))["title"])

# as a DataFrame
df = ds.to_pandas()
df["datePublished"] = pd.to_datetime(df["datePublished"], errors="coerce", utc=True, format="mixed")
recent = df[df["datePublished"].dt.year >= 2024]
print(len(recent), "articles from 2024 onwards")

Combining several outlets into one comparative corpus:

from datasets import load_dataset, concatenate_datasets

outlets = ["SinclairSchneider/zeit_de", "SinclairSchneider/tagesschau_de"]
parts = []
for name in outlets:
    d = load_dataset(name, split="train").select_columns(["title", "content", "url"])
    parts.append(d.add_column("outlet", [name.split("/")[-1]] * len(d)))
corpus = concatenate_datasets(parts)

Limitations and biases

• Crawl, not archive. Coverage is whatever was reachable by following links from the seed URL at crawl time. Articles that were deleted, depublished or orphaned before the crawl are missing, and old content is covered less densely than recent content.
• Free content only. Paywalled articles were deliberately skipped, so the dataset under-represents long-form and investigative pieces, which are the ones most often placed behind the paywall.
• Deduplication by headline. Two genuinely different articles with an identical headline (common for recurring formats such as weather or daily briefings) collapse into one row.
• Noisy metadata. author, category and keywords come straight from the publisher's markup, which changed over the years; empty strings are common and category labels are not harmonised across outlets.
• Editorial perspective. This is the output of a single outlet. Any model trained on it alone inherits that outlet's selection of topics, framing and vocabulary — which is exactly why the collection provides the same pipeline across the political spectrum.
• No text normalisation. Bodies are the concatenated paragraphs as published, including typographic quotes, soft hyphens and occasional residual boilerplate.

Licensing and intended use

Scientific research only — Forschungsprivileg (§ 60d UrhG)

This dataset is made available solely for the purpose of scientific research. The legal basis is the German text-and-data-mining exception for research, the so-called Forschungsprivileg:

• § 60d UrhG — Text und Data Mining für Zwecke der wissenschaftlichen Forschung — allows research organisations to make reproductions of lawfully accessible works in order to carry out text and data mining for non-commercial scientific research, and to share the resulting corpus with a specifically limited circle of people for joint research and with individual third parties for the purpose of quality assurance.
• It implements Art. 3 of Directive (EU) 2019/790 (the DSM Directive); comparable research exceptions exist in the other EU member states.
• § 60d UrhG does not cover commercial exploitation. Research organisations that give a commercial enterprise preferential access are explicitly excluded from the privilege.

What this means for you

✅ Permitted — using the corpus for non-commercial scientific research: text and data mining, corpus linguistics, media-bias and discourse analysis, training and evaluating research models, reproducing published findings, and quality checks by individual third parties.

❌ Not permitted — commercial use or commercial model training; redistributing or republishing the article texts; presenting the texts as your own; building products or services on the article texts; any use that substitutes for reading the publisher's own website.

The article texts remain the copyright of ZEIT and its publisher. No licence to them is granted or implied by this repository. Only the crawling and extraction code in scraping/ is offered for free reuse.

If your intended use falls outside the research exception, obtain permission from the rights holder, or use only the derived, non-expressive data (statistics, embeddings, counts). Note that § 60d UrhG also expects research corpora to be deleted once they are no longer needed for the research purpose, with archiving permitted at trusted institutions.

If you are a rights holder and want content removed, please open a discussion on this repository.

This section is an orientation for researchers, not legal advice. Responsibility for your own use of the data remains with you and your institution.

Citation

@misc{zeit_de_dataset,
  title        = {DIE ZEIT / ZEIT ONLINE (zeit.de): a corpus of German news articles},
  author       = {Schneider, Sinclair},
  year         = {2025},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/datasets/SinclairSchneider/zeit_de}
}

See also

• The full German Media collection
• Mirror of this dataset: NLP-UniBW/zeit_de
• medienkompass.org — Deutsche Medienlandschaft, the source of the perceived-partisanship figures above