mischeiwiller/german-job-postings

German Job Postings (Stellen-Atlas)

The open map of the German labour market's skills. A normalized corpus of 70,584 real German (DACH) job vacancies, each linked to the EU-official ESCO occupation/skill taxonomy (incl. the ESCO green-skills extension) and the German KldB-2010 occupation classification.

Open ML resources for the labour market are overwhelmingly US/English. This is, to our knowledge, the first open, in-language, ESCO-and-KldB-linked corpus of real (non-synthetic) German job postings. It is part of the Stellen-Atlas project (dataset + jobbert-de model + demo Space).

⚠️ Status: v1, taxonomy labels are machine-generated (zero-shot). The ESCO occupation/skill tags are bootstrapped with a zero-shot embedding model, not hand-verified. Quality is mixed — see Limitations. Treat the taxonomy fields as weak/bronze labels suitable for training and analysis, not as ground truth.

Dataset at a glance

Rows	70,584
Language	German (de)
Source	Bundesagentur für Arbeit (BA) Jobbörse API
Taxonomy	ESCO v1.2.0 (occupations + skills + green flag); KldB-2010 → ISCO-08 → ESCO
Rows with an ESCO occupation	70,580 (99.99%)
Rows with ≥1 ESCO skill	70,580; avg 4.98 skills/row
Rows flagged green (ESCO)	5,031 (7.1%)
Rows with a KldB-2010 code	70,448 (99.8%)
Rows with an observed salary range	1,254 (1.8%)
Format	Parquet (single file, ~4.5 MB)
License	CC-BY-4.0

Intended use

• HR-tech / job boards — skill tagging, occupation normalization, matching.
• Labour-market & policy analysis — green-transition skill demand, regional gaps (NUTS + ESCO green flag).
• German NLP / RAG — a clean, in-language jobs corpus.
• Research — ESCO/KldB-linked real postings; weak-label source for training skill-extraction models (see the companion mischeiwiller/jobbert-de).

Out of scope / non-goals

• No résumé/CV data. No salary prediction (only observed ranges, where stated). No non-DACH languages in v1. No verbatim re-publication of ToS-restricted board content (see Provenance & legal).

Schema

One row = one normalized vacancy.

Field	Type	Notes
id	string	Stable, unique within the corpus.
source	string	Origin system (bundesagentur).
source_url	string	Canonical URL of the original posting.
license	string	License governing the row's published fields (CC-BY-4.0).
provenance	string	Audit trail of how the row was obtained/derived.
fetched_at	timestamp (UTC)	When the row was retrieved.
posted_date	date	When the vacancy was posted.
lang	string	BCP-47 language code (de).
title	string	Job title.
employer	string | null	Hiring organization.
description_derived	string | null	Derived/cleaned description (ToS-safe). Empty in v1 — see below.
region	string | null	Region label (Bundesland/locality).
nuts_code	string | null	NUTS region code (e.g. DE21).
esco_occupation	struct{uri,label} | null	Mapped ESCO occupation.
esco_skills	list[struct{uri,label}]	Mapped ESCO skills.
is_green	bool	True if ≥1 mapped skill is in the ESCO green collection.
kldb_2010	string | null	German KldB-2010 occupation code.
seniority	string	Parsed seniority (entry/junior/mid/senior/lead/unknown).
salary_range	struct{min_eur,max_eur,period} | null	Observed range; never predicted.

A note on description_derived

The v1 corpus is built from the BA Jobbörse search/list endpoint, which returns structured fields (title, employer, location, KldB, dates) but not free-text ad bodies. To stay clearly within the source's terms of service we do not scrape and republish the full ad text, so description_derived is empty in v1. Taxonomy enrichment therefore runs over the title + occupation signal. A future version may add a ToS-safe derived summary.

How it was built

1. Ingest — official BA Jobbörse API over a 272-query (keyword × region) matrix plus country-wide sweeps → 151,957 raw records (cached, gitignored).
2. Normalize — raw → unified schema; region → NUTS/Bundesland; seniority and salary parsed from German text.
3. PII scrub + dedup — strip emails/phones/named contacts; exact dedup on (source, source_url) + near-dup on normalized title+employer+region. 151,957 → 70,584 rows.
4. Enrich (zero-shot) — intfloat/multilingual-e5-base embeds the German ESCO label space; nearest-neighbour assigns a candidate ESCO occupation and skills; the green flag and KldB code are resolved against the ESCO reference tables and the KldB-2010 ↔ ISCO-08 ↔ ESCO crosswalk.

Reproduction code: https://github.com/lappemic/hf-lab (project 01-stellen-atlas).

Provenance & legal

• Primary source: Bundesagentur für Arbeit (BA) Jobbörse API (https://jobsuche.api.bund.dev/), the official German public employment service. We publish derived/normalized structured fields, not verbatim ad bodies; PII (named contacts, emails, phone numbers) is scrubbed.
• ESCO v1.2.0 (occupations, skills, green-skills collection) — © European Union, CC-BY-4.0 (https://esco.ec.europa.eu/).
• KldB-2010 ↔ ISCO-08 crosswalk — Statistik der Bundesagentur für Arbeit, Umsteigeschlüssel KldB 2010 (Fassung 2020) → ISCO-08. Used only as reference data to derive the per-row kldb_2010 code — an eigene Berechnung (own calculation), marked as such; the verbatim crosswalk and the KldB/ISCO classification tables are never redistributed. The BA Statistik terms permit unrestricted use of the statistical tables the agency publishes under its statutory duty (no non-commercial clause), conditioned on not altering/falsifying the content and marking own calculations as such. © Statistik der Bundesagentur für Arbeit (https://statistik.arbeitsagentur.de/).
• KldB-2010 classification (reference) — Statistisches Bundesamt (Destatis), GENESIS-Online, where used, under Datenlizenz Deutschland – Namensnennung – Version 2.0 (dl-de/by-2-0) — attribution-only, commercial use permitted, no ShareAlike (one-way compatible into CC-BY-4.0). Quellenvermerk: "Datenquelle: Statistisches Bundesamt (Destatis), Genesis-Online; Datenlizenz by-2-0; eigene Berechnung."
• Dataset license: CC-BY-4.0. Attribute Stellen-Atlas / mischeiwiller and the upstream sources above (BA, Destatis, ESCO). Only derived codes — not the BA crosswalk or the KldB/ISCO classification tables — are distributed.

Limitations

• Machine-generated taxonomy labels (mixed quality). ESCO occupation/skill tags come from a zero-shot embedding model with no human review. They are reliable for clear, common titles but drift on terse or ambiguous titles (e.g. generic "Mitarbeiter (m/w/d)"). A spot-check on the development sample found strong top-1 occupation agreement at the ISCO major-group level for clear titles; tail skills carry noise. Use the taxonomy fields as weak labels, and prefer the forthcoming hand-checked gold split for evaluation.
• No ad body text (description_derived empty) — see the note above.
• Source bias. A single source (BA) and a keyword/region query matrix mean coverage skews toward the queried roles/regions; it is not a representative census of the German labour market. NUTS distribution is led by Bayern (DEA), Berlin (DE2), and Nordrhein-Westfalen (DEA/DE).
• Sparse salary. Only ~1.8% of rows state a salary range.
• is_green is taxonomy-derived, inheriting the zero-shot skill noise.
• No deduplication across re-postings of the same role by different refs is guaranteed beyond the title+employer+region near-dup heuristic.

Companion artifacts

Part of the Stellen-Atlas project (dataset + model + demo):

• Model: mischeiwiller/jobbert-de — German job-title → ESCO occupation classifier trained on this corpus.
• Space: mischeiwiller/stellen-atlas — paste a German ad → ESCO occupation + skills + green share.

Citation

@misc{stellen_atlas_2026,
  title        = {German Job Postings (Stellen-Atlas)},
  author       = {Scheiwiller, Michael},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  note         = {Derived from the Bundesagentur f\"ur Arbeit Jobb\"orse API;
                  linked to ESCO v1.2.0 and KldB-2010. CC-BY-4.0.}
}