You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

ResumeBench

ResumeBench

Quick links: [Website] [Code & CLI] [Careerflow]

Given a resume PDF and a 14-section JSON Schema, a system returns structured data that matches the source document. ResumeBench measures how accurately extraction systems perform this task across 300 curated real-world, handwritten, and adversarial resumes.

Resume parsers drive job matching, candidate screening, and career tools. Extraction errors propagate. ResumeBench measures accuracy at the field level and penalizes both omissions and hallucinations. Scoring is deterministic with no LLM judge.

The evaluation framework runs pipelines and reports scores by split, tag, and system. 20 documents are released publicly; to access the full benchmark (300 curated resumes), request access.

Dataset

ResumeBench contains 300 curated resumes (837 pages). Each document is tagged by source, difficulty, and layout traits.

Source Documents Pages Description
Real-world 138 367 Anonymized resumes via Careerflow
Handwritten 9 42 Handwritten resumes on paper
Synthetic 153 428 Expert-built (Canva, LaTeX, Overleaf) and generated resumes with deliberate parsing challenges
Total 300 837

The benchmark spans a range of difficulty challenges: multi-column and sidebar layouts, handwritten resumes, custom sections (bar admissions, military service, publications), adversarial text (name-skill collisions, contradictory dates), and scanned or rotated documents.

Document Gallery

Everyday resumes

Designed templates (Canva, LaTeX, Overleaf)

Specialized & adversarial

Leaderboard

300 curated resumes, 18 systems. Every system receives the identical PDF, the same 14-section JSON Schema, and the same extraction prompt. The F1 column is Field F1 (the primary metric), macro-averaged across sections per document, then across documents. Scores are percentages.

Rank System Type F1 Cost/page Latency
1 LlamaParse Extract Agentic Plus API 87.1572 $0.054 69.7s
2 Reducto Deep Extract API 85.5526 $0.040 64.8s
3 LlamaParse Extract Agentic API 85.5463 $0.023 54.6s
4 Claude Opus 4.8 LLM 85.5052 $0.090 31.3s
5 Gemini 3.5 Flash LLM 85.4518 $0.035 42.8s
6 GPT-5.6 LLM 84.7483 $0.060 38.5s
7 Extend Max Context API 84.7241 $0.030 45.2s
8 Qwen 3.8 Flash LLM 83.5181 $0.004 22.1s
9 Qwen 3.8 27B LLM 83.7681 $0.003 18.6s
10 LlamaParse Extract Cost-Effective API 83.5164 $0.009 32.4s
11 GPT-5.5 LLM 83.5022 $0.045 35.2s
12 Reducto Extract API 81.6959 $0.020 28.3s
13 Extend Extract API 80.9701 $0.015 33.1s
14 Qwen 3.5 35B LLM 77.9481 $0.002 15.8s
15 Qwen 3.5 9B LLM 75.7778 $0.001 10.2s
16 Gemma 4 26B LLM 62.8207 $0.003 19.4s
17 GPT-5.4 Nano LLM 61.8600 $0.008 12.5s
18 GLM 5.3 Flash LLM 58.8454 $0.002 14.3s

Per-section breakdowns, difficulty splits, and interactive filtering at resumebench.ai.

Metric details
  • F1: F1 of extracted values, macro-averaged across sections per document, then averaged across documents.
  • Identity-gated matching: Entities in structured sections are matched by identity fields (e.g. company + position + startYear + endYear for experience). All identity fields must match before the entity pair is scored. Keyword sections (languages, interests, profiles, publications) use Hungarian alignment.
  • String comparison: Whitespace collapsing. Case and punctuation count. Typographic variants (smart quotes, em dashes, soft hyphens) are folded to ASCII equivalents. Contacts and locations are compared case-insensitively.
  • Description bullets: Compared positionally (bullet[0] vs bullet[0], etc.) using normalized edit distance. Missing or extra bullets score zero.
  • Failed documents: Score zero (not dropped), so a system can't boost its average by erroring on hard ones.
  • Cost/page: USD per page at each provider's official public pricing.
  • Latency: Wall-clock seconds per resume (median).
  • Type: API = document extraction service (ingests schema natively); LLM = vision language model (receives rendered pages + prompt).

Scoring

Scoring is per-field with identity-gated matching:

  • Per-field units: Each field in a matched entity is scored independently. Description bullets are compared positionally (bullet[0] vs bullet[0], etc.).
  • Identity-gated matching: Entities are matched by identity fields (e.g. company + position + startYear + endYear). All identity fields must match before any other field is scored. Keyword sections (languages, interests, profiles, publications) and skills use Hungarian alignment.
  • Section-averaged: F1 is computed per section, then averaged. A missed custom section costs as much as a missed job.

Text comparison: strings are compared after whitespace collapsing. Typographic variants (smart quotes, curly dashes, soft hyphens) are folded to ASCII equivalents. Contacts and locations are compared case-insensitively. Failed extractions score 0, not skipped.

Usage

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

# Download the dataset
resume-bench download

# Run extraction
resume-bench run openai_gpt-5.6-sol_text --split test

# Grade results
resume-bench grade openai_gpt-5.6-sol_text --split test

# View leaderboard
resume-bench leaderboard --split test
Quick smoke test (mini split, 5 resumes)
resume-bench run openai_gpt-5.6-sol_text --split mini --limit 5
resume-bench grade openai_gpt-5.6-sol_text --split mini

Dataset Format

Splits

Split Documents Purpose
test 20 Public evaluation set (representative subset of the full 300)
mini 5 Quick smoke test (subset of test)

Record Schema

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": "John", "lname": "Doe", "email": "john@example.com"},
    "experience": {"sectionTitle": null, "items": [
      {"company": "Acme Corp", "position": "Engineer", "startYear": 2020, "endYear": 2023,
       "description": ["Led migration to microservices", "Reduced latency by 40%"]}
    ]},
    "education": {"sectionTitle": null, "items": [
      {"institution": "MIT", "area": "Computer Science", "studyType": "BS",
       "startYear": 2016, "endYear": 2020}
    ]},
    "skills": {"sectionTitle": null, "items": [
      {"category": "Programming", "skills": ["Python", "Go"]}
    ]},
    "personalSummary": {"sectionTitle": null, "text": "Experienced engineer with..."}
  },
  "difficulty": "medium",
  "layout_tags": ["real-anonymized", "table_grid"],
  "source": "real-anonymized",
  "domain": "software-engineering",
  "schema_version": "resume_v1"
}
14 scored sections
Section Type Identity Fields Description scored?
basics singleton (all scored fields) No
experience entity-list company + position + startYear + endYear Yes
education entity-list institution + studyType + startYear + endYear Yes
projects entity-list name + startYear + endYear Yes
personalSummary text text N/A
certifications entity-list name + issuer No
awards entity-list title No
volunteering entity-list organization + position + startYear + endYear Yes
skills flat-list (one unit per skill) No
publications keyword name (Hungarian) No
languages keyword name (Hungarian) No
interests keyword name (Hungarian) No
profiles keyword network (Hungarian) No
customSections entity-list sectionTitle Yes

Structured sections use direct greedy matching on identity fields (all must match). Keyword sections use Hungarian alignment.

Citation

@misc{kohli2026resumebench,
  title     = {ResumeBench: A Benchmark for Structured Resume Extraction},
  author    = {Puneet Kohli and Shiv Awasthi and Jeffy Sam},
  year      = {2026},
  url       = {https://resumebench.ai}
}

Links

Downloads last month
1,248