Datasetpublic

KSE-RESEARCH-Group/Work_UA_resumes

WorkUA Resumes Dataset Dataset Summary This dataset contains 103,895 structured resume entries collected from publicly available candidate profiles on Work.ua, Ukraine's largest job platform. Resumes were scraped, parsed, cleaned, and deduplicated for research use. Scraping window: July 9 – August 22, 2025. Intended use: Resume parsing and information extraction Ukrainian-language NLP pipelines Vacancy–candidate matching Labor market and salary analysis Career… See the full description on the dataset page: https://huggingface.co/datasets/KSE-RESEARCH-Group/Work_UA_resumes.

sourceHugging Facecc-by-nc-4.0updated 6mo agoView on Hugging Face

0likes42downloads

Dataset Card

WorkUA Resumes Dataset

Dataset Summary

This dataset contains 103,895 structured resume entries collected from publicly available candidate profiles on Work.ua, Ukraine's largest job platform. Resumes were scraped, parsed, cleaned, and deduplicated for research use.

Scraping window: July 9 – August 22, 2025.

Intended use:

—Resume parsing and information extraction
—Ukrainian-language NLP pipelines
—Vacancy–candidate matching
—Labor market and salary analysis
—Career recommendation systems
—Text classification and semantic search

Privacy: All direct personal identifiers have been removed — candidate full name, profile URL, and contact details are not present. Only the numeric resume id is retained for cross-referencing purposes.

Processing Pipeline

Resumes were scraped as HTML, converted to Markdown, then structured fields were extracted in three batches:

Batch	Resumes	Method	Description
Standard	84,245	Regex	Well-structured resumes with a consistent HTML layout. Fields like `work_experiences`, `educations`, `skills`, `languages`, etc. were extracted directly via regular expressions from the page DOM.
File-based	14,397	Gemini 2.5 Flash	Resumes uploaded by candidates as files (PDF, DOC, etc.). Work.ua renders a degraded text preview of these; the DOM-based regex approach was not viable. The full Markdown content was sent to Gemini with a strict JSON schema prompt (`temperature=0`) to extract all structured fields.
Extended	5,253	Gemini 2.5 Flash	Standard resumes where structured fields (`educations`, `work_experiences`, `skills`, `languages`, `driver_license`, etc.) were absent from the DOM but present inside the free-text `additional_info` blob. Gemini was used to parse these out and populate the same schema.

After concatenation and deduplication: 103,895 unique resumes.

File

`resumes.ndjson` — Newline-delimited JSON, 103,895 rows, 18 columns.

Schema

id                      String       — Numeric resume ID from Work.ua
title                   String       — Job title the candidate is seeking
age                     Int64        — Candidate age (27% null)
city                    String       — Candidate's city
desired_salary          Int64        — Monthly salary expectation in UAH (15% null)
employment_type         String       — e.g. "повна", "неповна" (47% null, inconsistently filled)
work_location_preference String      — e.g. "Дистанційно", "Офіс" (71% null, inconsistently filled)
driver_license          Boolean      — Whether a driver's license is mentioned
creation_date           String       — Date the resume was scraped (YYYY-MM-DD)
other_resumes           List[Struct] — Other resumes of the same candidate: {title, resume_id}
veteran                 Boolean      — Self-reported veteran status
disability              String       — Disability group: "Перша/Друга/Третя група" or null
work_experiences        List[Struct] — {position, start_date, end_date, company, city, industry, responsibilities}
languages               List[Struct] — {language, level}
skills                  List[String] — Free-text skill tags
educations              List[Struct] — {institution, faculty, city, level, start_year, end_year}
additional_educations   List[Struct] — Courses and certifications: {institution, start_year, end_year}
additional_info         String       — Free-text remainder that did not fit structured fields (61% null)

Data Example

json

{
  "id": "14656003",
  "title": "Розробник WordPress",
  "age": 24,
  "city": "Київ",
  "desired_salary": 37000,
  "employment_type": null,
  "work_location_preference": "Дистанційно",
  "driver_license": false,
  "creation_date": "2025-08-17",
  "other_resumes": [],
  "veteran": false,
  "disability": null,
  "work_experiences": [
    {
      "position": "Middle WordPress / Full-Stack Developer",
      "start_date": "2022-01-01",
      "end_date": "2025-01-01",
      "company": "CullyCully Studio",
      "city": "Швейцарія",
      "industry": null,
      "responsibilities": "Розробка комерційних сайтів та корпоративних сторінок..."
    }
  ],
  "languages": [
    {"language": "Українська", "level": "рідна"},
    {"language": "Англійська", "level": "середній"}
  ],
  "skills": ["WordPress", "WooCommerce", "PHP", "JavaScript"],
  "educations": [],
  "additional_educations": [],
  "additional_info": null
}

Known Limitations

Issue	Detail
High null rates in preference fields	`work_location_preference` (71% null) and `employment_type` (47% null) were inconsistently filled by candidates — do not treat null as "not specified"
Self-reported salary	Extreme values (e.g. 1 UAH, 400,000 UAH) are likely data entry errors; filter before use
Age anomalies	Max reported age of 125 — filter to a reasonable range before use
Inconsistent date formats	`work_experiences.start_date` / `end_date` vary: `"2022-01"`, `"2022-01-01"`, `"Present"`, `null`
Non-standardized free-text fields	`employment_type`, `languages.level`, and `skills` are raw free-text with no normalization
Extraction quality split	~81% regex (deterministic), ~14% Gemini for file resumes, ~5% Gemini for additional info extraction — the LLM-extracted portion may have subtle inconsistencies
`skills` noise	Mix of technical skills, soft skills, and personal traits with no taxonomy applied
Wartime context	43.7% of candidates self-reported veteran status — data reflects the Ukrainian labor market during the active war period (July–August 2025)

Ethical Considerations

—All resumes were scraped from publicly accessible pages on Work.ua
—No contact information (phone, email, address) is included
—Candidate full names and profile URLs were removed prior to publication
—Recommenders' names have been removed; the recommendations field is not included
—Do not attempt to re-identify individuals from this data
—Fields like veteran and disability reflect sensitive self-disclosed information and must be handled with care

Language

Primarily Ukrainian (uk). A minority of entries contain mixed Ukrainian/English text. No language filtering has been applied.

License

Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)

Free to use for non-commercial research with attribution.

Acknowledgments

—Data source: Work.ua
—AI extraction: Google Gemini 2.5 Flash
—Processing: Polars