PopResume

Population-Representative Resume Dataset for Causal Fairness Evaluation of LLM/VLM Resume Screeners

🌐 Project page · 📄 arXiv · 💻 Code · EMNLP 2026

PopResume is a synthetic, population-representative resume dataset that preserves the natural statistical relationships among demographic attributes, demographic proxies, and job-relevant qualifications found in U.S. population statistics. It enables path-specific effect (PSE)–based fairness auditing of LLM- and VLM-based resume screening systems — distinguishing legally permissible disparities (business necessity) from impermissible ones (redlining).

• Scale: 60,884 resumes across 5 occupations
• Occupations: accountants & auditors (7,784), construction laborers (5,910), elementary/middle school teachers (16,188), registered nurses (17,632), software developers (13,370)
• Formats: text resumes, resume images with / without a synthesized profile photo
• Rows: one per resume — 60,884 in text, 121,768 in images (× photo / no-photo)
• Protected attributes: sex (Female/Male); race (White, Asian/Pacific Islander, Black)
• Grounding: ACS PUMS, PSID, SSA name records, U.S. Census surname statistics

⚠️ Synthetic only. All resumes are generated by rule-based procedures from aggregate population statistics. No real individual's data is included. Raw IPUMS/PSID microdata record identifiers are not redistributed; only derived operational attributes are provided.

Following the paper, race == "AIAN" (American Indian / Alaska Native, 632 profiles) is excluded for stable causal estimation, giving 60,884 resumes. Rebuild with --include-aian for the full 61,516-resume superset.

Who is in this dataset?

Every profile is sampled per occupation from the 2023 ACS PUMS population of people who are currently employed in that occupation with a valid occupation code. Each synthetic candidate is therefore an incumbent of the job their resume targets — someone the labor market has already placed in that role. The candidate pool is qualified by construction: the benchmark measures how screeners score among qualified candidates, not whether they can filter out unqualified ones.

Configs

text — resume text + attributes

One row per resume id (60,884). Load:

from datasets import load_dataset
ds = load_dataset("sumin-yu/PopResume", "text")     # -> columns below

The three *_label / *_group name columns are the grouped redlining proxies the paper's estimator actually consumes — raw names are too high-cardinality for stable estimation. They are derived from public SSA and Census statistics by the thresholding rule given in the paper's appendix, and are reproduced by 1_data_generation/01-3_name_sampler.ipynb in the code repository.

Causal roles follow the paper's path-specific effect decomposition — X protected attribute, Z confounder, B business-necessity mediator, R redlining proxy. This is one operationalization motivated by legal and policy discussion; alternative groupings can be defined and audited the same way.

Column	Role	Source	Description
id	key	—	zero-padded resume id
job	—	ACS PUMS	occupation
sex, race	X	ACS PUMS	operational protected attributes
age, region	Z	ACS PUMS	confounders
edu_level, edu_group	B	ACS PUMS	education
exp_year	B	PSID	years of work experience
first_name, last_name	R	SSA, U.S. Census	name-based demographic proxies
first_name_sex	R	SSA	gender typicality of the first name: F / M / neutral (P ≥ 0.75)
first_name_age_group	R	SSA	age typicality: [17,25) / [25,35) / [35,45) / neutral (P ≥ 0.5)
surname_race_label	R	U.S. Census	race typicality of the surname: White / Black / Asian/Pacific Islander / neutral (P ≥ 0.5)
state	R / Z	ACS PUMS	address proxy
text	content	—	natural-language resume

images — resume images

Resume images in two variants — with a synthesized profile photo and without (2 per resume id, 121,768 images total). Every row pairs a decoded image with the same attributes as text, plus a has_photo flag.

ds = load_dataset("sumin-yu/PopResume", "images", split="train")
ds[0]["image"]        # PIL.Image
ds[0]["has_photo"]

Images are stored as sharded parquet (data/images/<occupation>-NNNN.parquet, ~450 MB per shard) with a native HF Image feature — so streaming=True works and you can pull a single occupation without downloading all ~20 GB:

ds = load_dataset("sumin-yu/PopResume", "images", split="train", streaming=True)
# or, one occupation only:
from datasets import load_dataset
ds = load_dataset("parquet", data_files="hf://datasets/sumin-yu/PopResume/data/images/registered_nurses-*.parquet")

Not included here

Model scores and causal-effect estimates are not part of this dataset — they are experiment outputs. This dataset ships only the benchmark inputs (resumes + attributes).

Resumes are rendered without a skills section, matching the configuration the paper evaluates.

License & terms

Released under CC BY 4.0. Derived from public U.S. statistical sources (ACS PUMS, PSID, SSA, U.S. Census); raw microdata are not redistributed here. Intended for research and practice on algorithmic fairness in automated hiring — including audits of deployed screening systems.

Citation

@inproceedings{yu2026popresume,
  title     = {PopResume: Causal Fairness Evaluation of LLM/VLM Resume Screeners
               with Population-Representative Dataset},
  author    = {Yu, Sumin and Park, Juhyeon and Moon, Taesup},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in
               Natural Language Processing (EMNLP)},
  year      = {2026}
}