README.md

---
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
---

# 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**: 38 documents (handwritten + adversarial distractors)
- **Schema Sections Scored**: 9 (basics, experience, education, projects, summary, certifications, awards, volunteering, skills)
- **Domains**: 6 (engineering, healthcare, legal, data science, government, general)
- **License**: CC-BY-4.0

**Quick Links**: [💻 Code & CLI](https://github.com/careerflow/resume-extract-bench) | [🏢 Careerflow](https://careerflow.ai)

---

## Dataset Introduction

### Document Composition

| Challenge Category | Documents | Description |
|---|---|---|
| **Handwritten resumes** | 28 | Scanned handwritten resumes testing OCR and layout understanding |
| **Adversarial distractors** | 10 | Synthetic resumes with intentional parsing challenges |
| **Total** | **38** | |

### Source Composition

- **Handwritten documents**: 28
- **Synthetic (LaTeX-generated)**: 10
- **All documents classified as "hard" difficulty**

### Challenge Categories

**C1: Handwritten Resumes** — Scanned handwritten documents with natural variation in handwriting, layout, and legibility. Tests OCR accuracy, spatial reasoning, and robustness to non-standard formatting. Typical failures: misread characters, merged/split words, missed sections, hallucinated content from ambiguous handwriting.

**C2: Adversarial Distractors** — Synthetic resumes with intentional parsing challenges designed to stress-test extraction robustness:

| Distractor | Challenge |
|---|---|
| **Date chaos** | Inconsistent date formats within the same resume (Jan 2020, 2020-06, 06/2021, Summer 2022) |
| **Esoteric titles** | Non-standard job titles (Chief Vibes Officer, Growth Hacker), salary listed, entity confusion |
| **European regional** | German-style CV with date of birth, marital status, nationality, military service, German grade scale |
| **Functional format** | Skills-first layout with no chronological work history |
| **Identity confusion** | Multi-part names, honorifics, maiden names, father's name, testimonial quotes with other people's titles |
| **Multilingual** | English resume with French and Spanish intertwined |
| **Prompt injection** | Hidden white text injection, PDF metadata injection, keyword stuffing in 6pt font |
| **Prose narrative** | Third-person writing, no bullet points, 500-word personal statement, emoji section markers |
| **Same company repeat** | Same employer listed 4 times with company renamed mid-tenure and overlapping dates |
| **Skill-name collision** | Candidate name collides with programming language (Ruby Chen at Python Solutions Inc.) |

### Domain Coverage

| Domain | Documents |
|---|---|
| Engineering | 11 |
| General | 10 |
| Healthcare | 7 |
| Legal | 7 |
| Data Science | 2 |
| Government | 1 |

---

## Usage

### Loading with Datasets

```python
from huggingface_hub import snapshot_download
import json

root = snapshot_download(
    repo_id="careerflow/ResumeExtractBench",
    repo_type="dataset",
)

with open(f"{root}/test.jsonl") as f:
    cases = [json.loads(line) for line in f if line.strip()]

schema = json.load(open(f"{root}/schema.json"))

for case in cases:
    pdf_path = f"{root}/{case['files']['pdf']}"
    ground_truth = case["ground_truth"]
    # Run your extractor on pdf_path against schema
    # Compare output to ground_truth
```

### Running Evaluation with CLI

```bash
pip install git+https://github.com/careerflow/resume-extract-bench.git
resume-bench download                                          # download dataset
resume-bench run --pipeline gpt-5.6 --split test               # run extraction
resume-bench grade --split test                                 # score against GT
resume-bench leaderboard                                        # view results
```

### Bring Your Own Predictions

```bash
resume-bench grade-file my_predictions.jsonl --split test
```

Predictions JSONL supports two formats:

**Flat format** (recommended):
```json
{"resume_id": "board-certified-ocularist-jane-doe", "basics": {"fname": "Jane", "lname": "Doe", ...}, "experience": [...], ...}
```

**Wrapped format**:
```json
{"resume_id": "board-certified-ocularist-jane-doe", "prediction": {"basics": {...}, "experience": [...], ...}}
```

### Dataset Files

- **test.jsonl**: 38 test cases (one JSON object per line)
- **schema.json**: Target JSON Schema for extraction (resume_v1)
- **pdfs/**: Source resume PDFs

---

## Dataset Format

Each JSONL line represents one resume test case:

```json
{
    "resume_id": "distractor-date-chaos-inconsistent-date-formats-throughout",
    "files": {"pdf": "pdfs/distractor-date-chaos-inconsistent-date-formats-throughout.pdf"},
    "ground_truth": { ... },
    "difficulty": "hard",
    "layout_tags": ["distractor", "inconsistent-date-formats"],
    "source": "synthetic-distractor",
    "domain": "software-engineering",
    "schema_version": "resume_v1"
}
```

### Field Definitions

| Field | Type | Description |
|---|---|---|
| `resume_id` | string | Unique identifier (PDF filename stem) |
| `files.pdf` | string | Relative path to source PDF |
| `ground_truth` | object | Human-verified structured extraction conforming to `schema.json` |
| `difficulty` | string | Difficulty level (`easy`, `medium`, `hard`) |
| `layout_tags` | list[string] | Visual and structural challenge tags |
| `source` | string | Resume origin: `expert-handwritten`, `synthetic-distractor` |
| `domain` | string | Professional domain of the resume |
| `schema_version` | string | Schema version (`resume_v1`) |

### Extraction Schema (resume_v1)

The schema defines 9 sections that must be extracted from each resume:

| Section | Type | Key Fields | Scored |
|---|---|---|---|
| `basics` | Singleton object | fname, lname, email, phone, city, state, country, hasPersonalPhoto | Per-field accuracy |
| `experience` | Entity list | company, position, startMonth/Year, endMonth/Year, city, description[] | Entity P/R/F1 + description token F1 |
| `education` | Entity list | institution, area, studyType, startMonth/Year, endMonth/Year, description[] | Entity P/R/F1 + description token F1 |
| `projects` | Entity list | name, url, description[] | Entity P/R/F1 + description token F1 |
| `personalSummary` | Free text | — | Token F1 |
| `certifications` | Entity list | name, issuer | Entity P/R/F1 |
| `awards` | Entity list | title | Entity P/R/F1 |
| `volunteering` | Entity list | organization, position, startYear, endYear, description[] | Entity P/R/F1 + description token F1 |
| `skills` | Flat list | category, skills[] | Set-level P/R/F1 |

See `schema.json` for the full JSON Schema definition.

---

## Evaluation Metrics

### Scoring Methods

1. **Entity-Level F1**: Entities (experience entries, education entries, etc.) are aligned between prediction and ground truth using the **Hungarian algorithm** (optimal bipartite matching via `scipy.optimize.linear_sum_assignment`). Match quality is determined by **Jaro-Winkler similarity** on key fields (threshold: 0.5).

2. **Description Token F1**: For matched entities with description arrays, a bag-of-words token F1 score measures description quality.

3. **Hallucination Rate**: Fraction of predicted entities that have no match in ground truth (spurious predictions).

4. **Omission Rate**: Fraction of ground truth entities that have no match in predictions (missed extractions).

All scoring is deterministic with no model in the loop.

### Aggregate Metrics

Per-section scores are averaged across all resumes. The **overall F1** is the mean of all section F1 scores, providing a single leaderboard ranking metric.

---

## Leaderboard

Results from the full 155-resume benchmark (includes this dataset plus additional medium-difficulty resumes):

| Rank | Model | Overall F1 | Halluc. Rate | Omission Rate |
|---|---|---|---|---|
| 1 | LlamaExtract Agentic Plus | **0.945** | 5.1% | 2.9% |
| 2 | Claude Opus | 0.937 | 5.1% | 3.6% |
| 3 | Extend Extract | 0.937 | 4.3% | 4.3% |
| 4 | Extend Max Context | 0.935 | 4.7% | 4.6% |
| 5 | Reducto Extract | 0.931 | 5.4% | 4.1% |
| 6 | Gemini 3.5 Flash | 0.928 | 5.8% | 4.3% |
| 7 | Reducto Deep | 0.923 | 6.3% | 4.9% |
| 8 | GPT-5.5 | 0.921 | 9.4% | 2.5% |
| 9 | GPT-5.6 | 0.920 | 8.7% | 3.2% |
| 10 | Gemma 4 26B | 0.914 | 6.5% | 5.4% |

*19 models benchmarked. Full results available in the [code repository](https://github.com/careerflow/resume-extract-bench).*

---

## Tag Taxonomy

ResumeExtractBench tags documents along three axes:

| Axis | Tags | Description |
|---|---|---|
| **Source** | `expert-handwritten`, `synthetic-distractor` | How the resume was created |
| **Difficulty** | `easy`, `medium`, `hard` | Overall extraction difficulty |
| **Domain** | `engineering`, `healthcare`, `legal`, `data-science`, `government`, `general` | Professional domain |

Layout tags provide additional visual/structural metadata per resume (e.g., `handwritten`, `fancy-templates`, `canva`).

---

## Citation

```bibtex
@misc{careerflow2026resumeextractbench,
    title={ResumeExtractBench: A Benchmark for Schema-Guided Resume Extraction},
    author={Careerflow and LlamaIndex},
    year={2026},
    url={https://huggingface.co/datasets/careerflow/ResumeExtractBench},
}
```

## License

All documents are synthetic or expert-created with fictional personal information. Released under [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/).