Datasets:

Careerflow
/

ResumeBench

license: cc-by-4.0
task_categories:
  - document-question-answering
  - feature-extraction
language:
  - en
tags:
  - structured-extraction
  - information-extraction
  - document-understanding
  - pdf
  - resume
  - json-schema
  - benchmark
size_categories:
  - n<1K
pretty_name: ResumeExtractBench
configs:
  - config_name: default
    data_files:
      - split: test
        path: test.jsonl
      - split: mini
        path: mini.jsonl

ResumeExtractBench

ResumeExtractBench is a benchmark for schema-guided structured extraction from resume documents. Given a resume PDF and a JSON Schema, systems must return structured data covering personal details, work history, education, skills, and more.

Dataset Size: 30 documents (real-world anonymized + synthetic + handwritten + expert-created)
Schema Sections Scored: 14 (basics, experience, education, projects, personalSummary, certifications, awards, volunteering, skills, publications, languages, interests, profiles, customSections)
Domains: 11 (engineering, software-engineering, legal, government, healthcare, finance, education, consulting, design, non-profit, general)
License: CC-BY-4.0

Quick Links: Code & CLI | Careerflow

Dataset Introduction

Document Composition

Source Category	Documents	Description
Real-world anonymized	5	Careerflow user resumes with PII replaced via synthetic anonymization
Synthetic forge	5	Generated resumes with diverse Word/LaTeX templates and real-like layouts
Expert-created (Canva)	5	Manually designed resumes using Canva templates with intentional distractors
Synthetic LaTeX	4	Template-diverse LaTeX resumes (academic, executive, ATS, creative styles)
Handwritten	4	Scanned handwritten resumes testing OCR and layout understanding
Synthetic legal	3	LaTeX-generated legal professional resumes with domain-specific content
Synthetic government	4	LaTeX-generated government/military resumes with clearances, MOS codes
Total	30

Domain Distribution

Domain	Count
Government	7
Legal	6
Engineering	4
General	3
Education	3
Finance	2
Consulting	2
Software Engineering	1
Non-profit	1
Design	1

Scoring Methodology: ResumeBench-Strict

This benchmark uses ResumeBench-Strict scoring with three rules that make it harder than per-cell scoring:

Record-level units. An entity (one job, school, certification, etc.) is ONE unit: it is correct only if it is aligned AND every non-empty field matches AND its bullets match one-for-one (same count, each bullet equal).
Empty is not free. Basics fields count only where the label or the parser has a value. Skills: one unit per skill.
Section-averaged. Per document, F1 / precision / recall are computed per section and averaged, so every section weighs the same.

Verbatim mode (on by default): strings match only after whitespace collapsing — case and punctuation count. Typographic variants (smart quotes, em dashes, soft hyphens) are folded.

Quick Start

git clone https://github.com/careerflow/resume-extract-bench.git
cd resume-extract-bench
pip install -e ".[dev]"

# Download the dataset
resume-bench download

# Run extraction (example: Anthropic Haiku, needs RESUME_BENCH_ANTHROPIC_API_KEY)
resume-bench run anthropic_claude-haiku-4-5_text --split mini --limit 5

# Grade results (verbatim by default)
resume-bench grade anthropic_claude-haiku-4-5_text --split mini

# View leaderboard
resume-bench leaderboard --split mini

Dataset Format

Splits

Split	Count	Purpose
`test`	30	Full evaluation set
`mini`	5	Quick smoke test

Record Schema (JSONL)

Each line in test.jsonl / mini.jsonl:

{
  "resume_id": "009b8ddc-4d4e-4452-ba2a-cc43ef83df77",
  "files": {"pdf": "pdfs/009b8ddc-4d4e-4452-ba2a-cc43ef83df77.pdf"},
  "ground_truth": {
    "basics": {"fname": "...", "lname": "...", "email": "...", ...},
    "experience": {"sectionTitle": null, "items": [{"company": "...", "position": "...", "description": [...]}]},
    "education": {"sectionTitle": null, "items": [{"institution": "...", "area": "..."}]},
    "skills": {"sectionTitle": null, "items": [{"category": "...", "skills": ["Python", "Go"]}]},
    "personalSummary": {"sectionTitle": null, "text": "..."},
    ...
  },
  "difficulty": "medium",
  "layout_tags": ["real-anonymized", "table_grid"],
  "source": "real-anonymized",
  "domain": "software-engineering",
  "schema_version": "resume_v1"
}

14 Scored Sections

Section	Type	Key Fields
`basics`	singleton	prefix, fname, lname, suffix, email, phone, city, state, country, linkedinUrl
`experience`	entity-list	company, position, dates, city/state/country, description bullets
`education`	entity-list	institution, area, studyType, score, dates, descriptions
`projects`	entity-list	name, dates, url, description
`personalSummary`	text	Free-text professional summary
`certifications`	entity-list	name, issuer, date, url
`awards`	entity-list	title, awarder, date, summary
`volunteering`	entity-list	organization, position, dates, description
`skills`	flat-list	Flat skill names across categories
`publications`	entity-list	name, publisher, date, summary
`languages`	entity-list	name, level
`interests`	entity-list	name
`profiles`	entity-list	network, url
`customSections`	entity-list	sectionTitle, summary, description

Grading Your Own Model

Python API

from resume_bench.grading.grader import grade_single_strict

result = grade_single_strict(ground_truth, prediction, verbatim=True)
print(f"F1: {result['f1']:.4f}")
print(f"Sections: {result['sections']}")

JSONL Predictions File

# Create predictions file
echo '{"resume_id": "abc123", "prediction": {...}}' > preds.jsonl

# Grade
resume-bench grade-file preds.jsonl --split test

Citation

@misc{resumeextractbench2026,
  title={ResumeExtractBench: A Benchmark for Schema-Guided Structured Extraction from Resumes},
  author={Careerflow and LlamaIndex},
  year={2026},
  url={https://huggingface.co/datasets/careerflow/ResumeExtractBench}
}