# Datasheet — Resume Parsing (Vision, Synthetic)

Following Gebru et al., "Datasheets for Datasets". Figures describe the v1.0 release
(1000 samples; target 1000).

## Motivation

- **Purpose.** A privacy-safe, openly redistributable dataset for training and evaluating
  resume-parsing models that read a resume image and emit a structured 23-field record. It is the
  companion dataset to the `sukhrobnurali/qwen3vl-resume-parser` model, created because the original training corpus was
  private third-party personal data that cannot be shared.
- **Created by.** sukhrobnurali.

## Composition

- **Instances.** Each instance is one fictional resume: 1-3 rendered page images, the canonical
  23-field JSON record, and the rendered plain text. 1000 instances in this release.
- **Sampling.** Personas are sampled from a seeded generator across 32 software-industry
  roles, 0-25 years of experience (mean 7.9),
  several countries, and varied section completeness. Visual diversity: 3 layouts x 4 themes.
- **Labels.** The JSON record is authored by code (not annotated after the fact), so labels are
  exact. The rendered document is a faithful surface form of that record, verified by a
  text-fidelity QC gate.
- **No personal data.** All identities, employers, and institutions are fictional; contacts use
  reserved test domains/number blocks. The data is not derived from real individuals.
- **Splits.** train 850 / validation 75 / test 75, with no
  fictional employer spanning two splits.

## Collection / Generation Process

1. A seeded spec sampler fixes each persona's classified attributes (role, seniority, dates,
   countries, salaries, skills, languages, layout, theme).
2. A language model writes only the free-text prose and fictional identity strings for that spec.
3. Code assembles the canonical record, renders it to HTML and page PNGs, and runs QC gates
   (schema, enums, text fidelity, persona de-duplication, rendered page-count). Failures are
   regenerated; only fully passing samples are accepted and recorded in a ledger.

## Preprocessing / Cleaning / Labeling

No post-hoc cleaning: labels are the authored canonical record. Cross-sample persona
de-duplication and an employer-overlap-free split assignment are applied during generation.

## Uses

- **Intended.** Training/evaluating image-to-structured-JSON resume parsers; document-AI and
  information-extraction research; a privacy-safe benchmark substitute.
- **Out of scope / caution.** Not representative of the full diversity of real resumes (IT/software
  focus, English prose, generator conventions for dates/phones). Not for inferring anything about
  real people — there are none.

## Distribution

- Hugging Face dataset repo `sukhrobnurali/resume-parsing-vision`, parquet via the `datasets` library, license CC-BY-4.0.

## Maintenance

- Maintained by sukhrobnurali; grows toward 1000 samples. Splits are frozen so released
  samples keep their split across versions. Issues/PRs via the HF repo.