JobHop v2: A Large-Scale Career Trajectory Dataset from Unstructured Resumes

Technical Summary: JobHop v2

Problem Statement
Large-scale, richly annotated career trajectory data is essential for workforce planning, job recommendation systems, and labor market analysis. However, the field lacks a suitable public benchmark. Existing public datasets are either small (e.g., 2,164 or 301 resumes), derived from pre-standardized occupational codes rather than raw text, or built from proprietary platforms inaccessible for independent research. Furthermore, previous attempts to extract data from unstructured resumes often suffer from degraded quality on noisy, multilingual inputs and fail to produce consistently machine-parseable output.

Methodology
The authors present JobHop v2, a dataset constructed through an end-to-end Large Language Model (LLM) extraction pipeline from a corpus of approximately 440,000 pseudonymized, multilingual resumes provided by VDAB (the Flemish Public Employment Service).

• Data Source & Ethics: The raw data consists of resumes in Dutch, French, and English. To ensure privacy, personally identifiable information (PII) and low-frequency terms were replaced with <MASK> tokens by VDAB. All processing occurred on local infrastructure (two NVIDIA H200 GPUs) to prevent data transmission to external services. The final dataset removes location fields and coarsens all dates to quarter-level granularity.
• Extraction Pipeline: The core innovation is a reasoning-controlled LLM inference pipeline using the openai/gpt-oss-120b model.
  • Prompting: A strict, rule-rich instruction prompt guides the model to extract structured JSON without hallucination. It handles specific corpus challenges such as <MASK> tokens, multilingual content, inconsistent date formats, and multi-column layouts.
  • Retry Mechanism: A persistent retry loop reprocesses samples that fail JSON validation using an extended generation budget and increased repetition penalty. This ensures a 100% JSON parse rate across the ~400,000 input resumes (recovering ~12.6% of initial failures).
  • Schema: The output schema is significantly richer than the original JobHop, capturing work experiences (with ESCO-aligned standardized titles, contract types, and technical skills), education (mapped to a five-level taxonomy), languages, and certificates.
• Cleaning & Normalization: A five-stage cleaning pipeline follows extraction, addressing parsing artifacts, invalid date ranges, exact duplicates, and quality filtering. A subsequent normalization pipeline maps job titles to ESCO codes (using a two-step hybrid policy with confidence thresholds), converts dates to quarter-year granularity, and maps education to a standard degree-level taxonomy.
• Evaluation Protocol: The authors introduce a revised evaluation protocol using 200 benchmark resumes. Instead of binary scoring, they employ graduated partial-credit scoring, length-ratio-gated string comparison, and field-importance weighting. To disentangle extraction errors from annotation noise, performance is measured against three complementary baselines: original hand annotations, AI-edited labels, and independent AI-generated labels.

Key Contributions

1. JobHop v2 Dataset: A public release of 355,315 unique career trajectories containing 1.99 million work experiences and 923,981 education entries. It is distinguished by being the first large-scale public dataset to pair ESCO occupational codes with temporal information and normalized education levels, all extracted end-to-end from real, unstructured resumes.
2. Robust Extraction Pipeline: A reasoning-controlled LLM pipeline with a multi-step retry mechanism that achieves a 100% JSON parse rate. The best-performing model (GPT-OSS-120B with high reasoning effort) comes within 1.1–2.7 percentage points of the inter-annotator agreement ceiling across all three baselines.
3. Revised Evaluation Protocol: A scoring system that uses three annotation baselines to provide a more faithful assessment of extraction quality, distinguishing between genuine model errors and inherent annotation ambiguity.

Results

• Extraction Quality: In the evaluation against the three baselines, GPT-OSS-120B (high reasoning) achieved the highest scores, trailing the agreement ceiling by only 1.1–2.7 percentage points. This suggests that residual errors are largely due to annotation ambiguity rather than systematic model failure.
• Comparison with JobHop v1: In a blind pairwise comparison of 1,000 resumes judged by an LLM (Kimi K2.6), JobHop v2 was preferred over the original JobHop pipeline on 68.3% of resumes (vs. 29.9% for v1). The improvement was most significant in work experience extraction, with mean accuracy scores of 8.60 for v2 versus 7.81 for v1.
• Dataset Statistics: The dataset covers a broad occupational spectrum, with Professionals, Technicians, and Service & Sales workers being the most frequent. Career trajectory lengths are right-skewed, with a median of 5 jobs per resume.

Significance
The paper claims that JobHop v2 fills a critical gap in the field by providing a large-scale, public benchmark derived from authentic, unstructured text rather than synthesized or pre-standardized data. The dataset and its accompanying extraction pipeline are intended to support reproducible research in career-trajectory analysis, next-occupation prediction, and labor market modeling. The authors emphasize that the dataset preserves the real-resume foundation of the original release while substantially improving extraction quality, annotation breadth, and evaluation rigor. They also note that while ESCO standardization provides a neutral vocabulary, the dataset retains the structural inequities of the underlying labor market, necessitating rigorous fairness evaluation in future deployments.