STEP & JobHop v2

This repository contains the code for two related research projects on modelling career trajectories from unstructured resumes:

• jobhop/ — JobHop v2: an LLM-based information-extraction pipeline that turns raw, multilingual resumes into ESCO-coded career trajectories, plus the dataset-construction and evaluation code behind the JobHop v2 dataset.
• step/ — STEP: a career-path recommendation model (Sequential Trajectory of Employment Prediction) built on ROUTE occupational embeddings (Representation Of Unique career Trajectories as Embeddings). Includes embedding fine-tuning, the sequence predictor, baselines, and evaluation.

The two projects share the JobHop v2 dataset: jobhop/ builds it, and step/ uses it (alongside other public benchmarks) to train and evaluate recommendation models.

Papers

• STEP: https://arxiv.org/abs/2607.11722
• JobHop: https://arxiv.org/abs/2607.11715

The JobHop dataset

The released JobHop dataset is publicly available on the Hugging Face Hub:

https://huggingface.co/datasets/aida-ugent/JobHop

Note on data. The released JobHop / JobHop v2 dataset is public at the link above. The source resume corpus it was extracted from is private (pseudonymized HR data provided by VDAB, the Flemish Public Employment Service, under a research collaboration agreement) and is not included here. Scripts that operate on the raw corpus therefore expect you to supply your own data via environment variables (see each subproject's README). All example/fixture data in this repository is synthetic. The modelling code in step/ can be run end-to-end on the public benchmarks (Decorte, KARRIEREWEGE, JobHop) without the private corpus.

Layout

step/      STEP career-path recommendation (ROUTE embeddings + STEP predictor, baselines, eval)
jobhop/    JobHop v2 information-extraction pipeline, dataset construction, and evaluation

Each subproject has its own README with entry points and usage.

Installation

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Most scripts expect their subproject directory on the PYTHONPATH (the provided run scripts set this for you), e.g. export PYTHONPATH="$PYTHONPATH:$(pwd)/step".

License

Released under the MIT License — see LICENSE.