sukhrobnurali/resume-parsing-vision

Resume Parsing (Vision, Synthetic)

A fully synthetic vision dataset for resume parsing: rendered resume page images paired with the ground-truth structured JSON they encode. Companion to the sukhrobnurali/qwen3vl-resume-parser model.

Release v1.0: the complete 1000-sample dataset. Splits are frozen (see below), so any future additions never move an existing sample between splits.

What each sample contains

Field	Description
images	The 1-3 rendered resume page PNGs (the model input).
ground_truth	The 23-field canonical resume record as a JSON string (the parse target).
resume_text	The rendered resume as plain text (for text-only baselines).
sample_id, desired_position, years_experience, num_pages, template_id, theme_id, languages, split	Flat metadata for browsing and filtering.

The 23 ground-truth fields: first_name, last_name, date_of_birth, email, phone, desired_position, about, job_experience, job_expectations, min_salary, max_salary, ready_to_relocation, work_modes, employment_types, employment_durations, hobbies, address, skills, experiences, educations, languages, certificates, projects.

How it was generated

Each resume is built structured-first: code authors every classified/canonical value (role, dates, countries, salaries, skills, languages, contact details) and its surface form. A language model writes only free-text prose and fictional identity strings (names, company and institution names, summaries, descriptions) — it never invents the classified fields. The record is rendered to HTML, then to page PNGs across 3 layouts x 4 themes, and passed through automated QC gates (JSON-schema, enum, text-fidelity, persona de-duplication, and rendered-page-count). Only samples that pass every gate are kept. The canonical JSON is therefore authored, never parsed back from a document, so labels are exact by construction.

Splits

Frozen, employer-overlap-free splits at the final 85%/8%/8% proportions:

Split	Samples
train	850
validation	75
test	75

Samples linked by a shared (fictional) employer are kept together in one split, so no employer spans two splits. Each sample's split is fixed in the generation ledger; future releases only assign newly added samples, so a sample published here keeps its split forever.

Privacy: no real personal data

This dataset contains no real personal data. Every name, company, and institution is fictional; e-mail addresses use the RFC 2606 reserved domain example.com; phone numbers use the reserved 555-01xx fictional block. It is not derived from any real candidate corpus. It is intended as a privacy-safe, redistributable substitute for resume-parsing work.

Composition (this release)

• 1000 resumes spanning 32 software-industry roles.
• Experience: 0-25 years (mean 7.9).
• Pages per resume: 1pg x207, 2pg x788, 3pg x5.
• Optional sections are present at realistic rates (e.g. about 73%, projects 64%, certificates 39%, hobbies 35%).

Roles (top 12)

Role	Count
System Administrator	48
Support Specialist	41
Networking Specialist	38
Data Scientist	37
Mobile Developer(IOS/Android)	37
Scrum Master	36
Security Specialist	34
Developer	34
DB Administrator	34
Architect	34
Product Owner	34
Machine Learning Engineer	34

Languages

Language	Resumes
English	1000
Hebrew	271
Uzbek	265
Spanish	265
Russian	259
German	253
French	232

Limitations

• Scope: all 1000 planned samples are included.
• Domain: IT / software-industry roles only; prose is in English.
• A small number of given-name repeats exist among the first 25 samples (generated before a name-diversity safeguard was added); later samples are de-clustered.
• Phone numbers use a single fictional format; date and country surface forms follow the generator's conventions. See the datasheet for the full list.

License & citation

Released under CC-BY-4.0. If you use this dataset, please cite it and link back to this repo and the companion model sukhrobnurali/qwen3vl-resume-parser.

See DATASHEET.md for the full datasheet (Gebru et al.).