JobHop

A large-scale public dataset of career trajectories derived from pseudonymized resumes provided by VDAB, the public employment service in Flanders, Belgium. Job experiences are standardized to ESCO occupation codes and carry quarter-level temporal information.

Two versions are available: JobHop v1 (the original release, ~360,000 resumes) and JobHop v2 (the current release, 355,315 trajectories from a larger corpus, built with a substantially improved extraction pipeline). Details of each are below.

📝 Dataset Description

JobHop v1

Built from approximately 360,000 resumes. Information was extracted with a small language model, and each job experience was mapped to an ESCO code using a proprietary tool provided by Nobl.ai. Documented in the JobHop paper, published at IEEE BigData 2025.

Important remark: On July 2nd we uploaded a cleaned version of the data. The cleaning primarily involves changing tags of ambiguous job titles to unknown and removing some rows.

JobHop v2

JobHop v2 is built from a corpus of approximately 440,000 pseudonymized VDAB resumes. This corpus is predominantly Dutch (91.4%), with smaller portions in English (6.1%) and French (2.4%), as reported for the original JobHop release. After filtering out documents left empty or near-empty by pseudonymization, about 400,000 resumes entered extraction. The release comprises 355,315 career trajectories, 1,993,291 work-experience entries, and 923,981 education entries.

Relative to v1, the pipeline was redesigned end to end:

• Extraction uses openai/gpt-oss-120b at high reasoning effort, run entirely on internal infrastructure as required by the data agreement, with a retry mechanism that achieves a 100% JSON parse rate.
• A richer extraction schema, a five-stage cleaning workflow, and normalization of dates to quarter granularity and education to a five-level scale.
• ESCO assignment (taxonomy v1.1.2) uses a two-step policy over the Nobl.ai classifier: a primary pass on title + description, and a stricter high-precision rescue pass on the model-generated standardized title. Entries unassigned after both steps are labeled unknown.
• In a blind pairwise comparison over 1,000 resumes with the original resume as reference, an LLM judge preferred v2 extractions on 68.3% of resumes versus 29.9% for v1.

Documented in "JobHop v2: A Large-Scale Career Trajectory Dataset from Unstructured Resumes" (RecSys in HR 2026 workshop; https://arxiv.org/abs/2607.11715).

📊 Dataset Structure

Field	Description
person_id	Identifier indicating which resume the row belongs to
esco_code	Matched ESCO code for the extracted experience (unknown where no confident match was found)
start_date, end_date	Quarter-level dates for the job experience (Q# YYYY)
university_degree	v1 only. Binary flag indicating a university degree
education_level	v2 only. Normalized highest education attainment on a five-level scale (None, Secondary, Bachelor, Master, PhD)

In v2, education_level replaces the binary university_degree flag; all other fields are unchanged from v1.

What is not in the release. The v2 extraction pipeline uses a richer schema than the published fields — it also extracts standardized job titles, free-text descriptions, company and institution names, technical skills, languages, certificates, contract type, and work-schedule type. Those fields are used internally to produce the ESCO codes and education levels; the public release contains only the structured fields listed in the table above.

✅ Intended Uses

This dataset is released to support reproducible research on career trajectories. Suitable uses include:

• Next-occupation and career-path prediction
• Learning and evaluating occupational representations and embeddings
• Job and career-transition recommendation research
• Aggregate labour-market and mobility analysis
• Benchmarking LLM-based information extraction from resumes

🚫 Discouraged and Out-of-Scope Uses

• Do not use models trained on this data as the sole or decisive basis for high-stakes decisions about individuals — hiring, promotion, dismissal, benefit eligibility, or the allocation of training and employment services. Trajectory-based predictions reflect historical patterns of labour-market access, not individual merit or potential.
• Do not attempt re-identification of individuals, or linkage of these records against other datasets for that purpose. We ask all users to respect this, and the underlying research agreement with VDAB requires it.
• Do not treat the data as a representative sample of any national or regional workforce (see Limitations).
• Do not infer or impute protected attributes (gender, ethnicity, age, nationality, disability) from occupations, education, or career timing.
• Any deployed system built on this data should be accompanied by a fairness evaluation across relevant groups before use.

🔒 Privacy and Residual Risk

Several layers of protection apply to this release:

• VDAB pseudonymized all resumes before sharing, replacing personally identifiable information and low-frequency terms with <MASK> tokens.
• All processing was performed on internal infrastructure; resume contents were never transmitted to external services or third-party APIs.
• Location fields are removed from the release, and all dates are coarsened to quarter-level granularity.
• Free-text fields (job titles, descriptions, company names, institution names) are not part of the public release; only the structured fields listed in the Dataset Structure table are published.

These measures remove direct identifiers and substantially reduce re-identification risk, but they do not eliminate it. Rare combinations of occupations, education, and timing may remain distinctive. The release is therefore best characterized as strongly pseudonymized rather than fully anonymous. No formal residual linkage-risk assessment has been conducted. Users are asked not to attempt re-identification and to apply appropriate safeguards when redistributing derived data.

⚠️ Limitations and Representational Scope

• Geographic and linguistic scope. All resumes come from jobseekers registered with VDAB in Flanders, Belgium, and are predominantly Dutch. Generalization to other labour markets, languages, or institutional contexts is untested.
• Population. The corpus consists of people who registered with a public employment service and uploaded a resume. It over-represents jobseekers and under-represents people who never used the service, and it should not be read as a representative sample of the Flemish or Belgian workforce.
• Occupational composition. Service and sales workers (ISCO 5), professionals (ISCO 2), and technicians and associate professionals (ISCO 3) dominate; primary-sector occupations (ISCO 6) are rare.
• Extraction is imperfect. Fields are produced by an LLM from unstructured text. On a 200-resume benchmark the extractor scores within 1.1–2.7 percentage points of the agreement level observed between independent annotation sets — close, but not error-free. The most common sources of disagreement we observed between independent annotations were date granularity, conventions for ambiguous section boundaries, how much free-text detail is copied into a description, and other subjective judgment calls.
• ESCO assignment. The 6.9% unknown rate is a coverage limitation. Confidence thresholds were set by manual inspection, and manual review of several hundred assignments found the labels to be of consistently high quality, but no systematic quantitative audit of ESCO label accuracy has been published.
• Reproducibility. Extraction, cleaning, normalization, and the evaluation protocol are released as open code at github.com/aida-ugent/Step, and rely only on openly available models. ESCO code assignment is the one proprietary step and cannot be reproduced without access to the Nobl.ai classifier; the assigned codes are, however, distributed with the dataset, so downstream users can work with the labels without re-running that step.
• Historical bias. The trajectories record real labour-market outcomes and therefore encode existing structural inequities in access, occupational segregation, and hiring. ESCO standardization provides a neutral vocabulary but does not remove those biases from the underlying data.

📜 Citation

To cite JobHop v1:

@inproceedings{jobhop-v1,
      title={JobHop: A Large-Scale Dataset of Career Trajectories},
      author={Johary, Iman and Romero, Rapha\"el and Mara, Alexandru C. and De Bie, Tijl},
      booktitle={2025 IEEE International Conference on Big Data (BigData)},
      pages={2184--2191},
      year={2025},
      doi={10.1109/BigData66926.2025.11402454},
      url={https://arxiv.org/abs/2505.07653},
}

Published at IEEE BigData 2025; the arXiv version is more complete and is the recommended reading.

To cite JobHop v2:

@inproceedings{johary2026jobhopv2,
      title={JobHop v2: A Large-Scale Career Trajectory Dataset from Unstructured Resumes},
      author={Iman Johary and Guillaume Bied and Alexandru C. Mara and Tijl De Bie},
      booktitle={Proceedings of the 6th Workshop on Recommender Systems for Human Resources (RecSys in HR 2026), co-located with the 20th ACM Conference on Recommender Systems},
      series={CEUR Workshop Proceedings},
      year={2026},
}

✍️ Authors

• Curated by: Iman Johary, Guillaume Bied, Alexandru C. Mara, Tijl De Bie (v2); Iman Johary, Raphaël Romero, Alexandru C. Mara, Tijl De Bie (v1)
• Funded by: BOF of Ghent University (BOF20/IBF/117), Flemish Government (AI Research Program), FWO (11J2322N, G0F9816N, 3G042220, G073924N), ERC grant (VIGILIA, 101142229)
• Data provided by: VDAB, the Flemish Public Employment Service
• License: CC BY 4.0

📧 Contact

Corresponding author: iman.johary@ugent.be