akzaidan/People

Synthetic US Candidate Profiles

248,522 synthetic job-seeker profiles covering 803 occupations across the United States, generated with DeepSeek V4 Flash. Built to train a job-ranking cross-encoder, where real candidate data cannot be used.

No real people. Every profile is generated. No resumes were scraped, no user records were used, and nothing here maps to a real individual. Names, emails, phone numbers and URLs are excluded by construction and checked for on every row.

Quick start

from datasets import load_dataset
import json

ds = load_dataset("<your-org>/<your-dataset>", split="train")
doc = json.loads(ds[0]["profile_json"])

print(doc["user"]["career_interests"])   # ['Recommendation Systems Engineer', ...]
print(doc["user"]["skills"][:5])         # ['Learning-to-rank', 'Feature stores', ...]
print(doc["experience"][0]["title"])     # 'Senior Machine Learning Engineer'

Schema

Seven flat columns for filtering without parsing JSON, plus the full record.

column	type	description
profile_id	string	unique identifier; prefix marks the generation batch (see Provenance)
source	string	always synthetic
cohort	string	generation cohort (13 values, see below)
job_family	string	one of 14 families
seniority	string	intern … executive
country	string	always United States
seed_role	string	the occupation the profile was seeded from (803 values)
profile_json	string	complete profile as a JSON object

profile_json is a single application-user document with exactly eight top-level keys:

key	contents
user	age, ethnicity, legal_status, sponsorship_needed, city, state, country, desired_salary, experience_years, skills, languages, bio, career_interests
work_location_preferences	where they want to work — remote / country / state / city entries, containment-checked and capped at 20
experience	company, title, start_date, end_date, description, city, state, country, is_internship
education	school, degree, field_of_study, start_date, end_date, is_current, gpa, description
projects	name, description, start_date, end_date
certifications	name, description, date
awards, publications	always empty — no source data

Dates are MM/YYYY strings; a null end_date on an experience entry means the role is current. state is a two-letter code ("IL") and country is ISO-2 ("US").

Four fields are computed or sampled rather than generated, and should be read that way:

field	how it is produced	why
user.experience_years	computed from experience date spans	LLMs are unreliable at date arithmetic, and this value feeds a scoring feature directly
user.age	derived from the earliest education/experience year	not generated, not real; a plausible value consistent with the work history
user.ethnicity	sampled from a census-like distribution, seeded by profile_id	present for coverage only — it is not inferred from the profile and carries no relationship to its content (see Limitations)
user.legal_status, sponsorship_needed	sampled from a fixed distribution	visa status is not reliably inferable from a résumé, and guessing it in a hiring context would be both inaccurate and wrong

Composition

Cohorts — profiles are deliberately not all well-formed. Real signups are often sparse, mid-career-change, or hard to classify.

cohort	n		cohort	n
standard	49,682		over_qualified	14,945
entry_level	32,267		visa_constrained	12,460
career_changer	24,785		return_to_work	12,444
sparse_profile	22,402		bootcamp_self_taught	7,511
senior_specialist	22,261		military_transition	7,446
student_intern	19,955		gig_contract	4,999
manager_lead	17,365

Seniority — senior 29.7%, entry 21.2%, mid 20.4%, junior 8.3%, intern 4.9%, manager 4.4%, staff 3.7%, principal 3.5%, director 2.6%, executive 1.4%.

Job families (14) — other 19.1%, operations 17.9%, software_engineering 15.7%, healthcare 10.4%, research 6.3%, hardware_engineering 6.0%, finance 4.9%, education 3.9%, design 3.7%, marketing 3.3%, legal 3.2%, sales 2.6%, data_science 2.1%, product_management 1.1%.

Occupations — 803 seed roles, each with 308–311 profiles. Coverage is even by construction rather than popularity-weighted, so welders, paralegals and nurses are represented on the same footing as software engineers.

Geography — 50 states and territories. California 23.0%, Texas 11.0%, Florida 6.0%, Colorado 4.0%, North Carolina 3.6%, Arizona 3.6%, Virginia 2.7%, Illinois 2.6%.

Content — median 13 skills (p10 8, p90 17), 3 experience entries (p10 1, p90 5), 10 years experience (p10 2, p90 20), age 33 (p10 25, p90 43), target salary $85,000 (p10 $48,000, p90 $145,000). 87.5% have a bio, 94.9% education, 79.6% certifications, 77.7% projects. 47.9% list Remote among their work-location preferences; 5.1% need sponsorship. Languages: English 99.9%, Spanish 19.9%, then a long tail.

Sampled demographics — ethnicity: White 57.5%, Hispanic/Latino 19.0%, Black 13.0%, Asian 6.3%, Other 1.7%, Native American 1.3%, Middle Eastern 1.2%. Legal status: Citizen 94.4%, Work Visa 2.7%, Student Visa 2.4%, Permanent Resident 0.5%. Both are drawn from fixed distributions, independent of profile content.

How it was built

Roles were enumerated systematically rather than sampled, so coverage is even rather than popularity-weighted. A vocabulary bank was generated once per occupation (803 roles → ~33,700 distinct skill, tool, specialization and certification terms), then every profile was seeded with a role, cohort, city, seniority, employer archetype, education path, career shape and a random subset of that role's vocabulary before generation at temperature 1.0. Seeding is positional rather than random, so each attribute is visited in proportion within every role instead of by luck.

Generation produces a richer internal record — used by the validators and the work-location assigner — which is converted to the stored document before writing; the internal shape is never persisted. Every profile was schema-validated at write time and again at merge, near-duplicates were dropped by per-role SimHash over the rendered text, and anything matching a PII pattern was discarded.

The 148,527 profiles in the newer batch cost $121.85 and about 4.5 hours at 96-way concurrency (measured, at off-peak API rates). 40 rows were dropped at merge: 37 near-duplicates, 2 schema failures, 1 possible-PII match.

Provenance. The corpus was built in two batches. syn_-prefixed IDs (99,995 rows) come from an earlier run that predates explicit seniority seeding; syn2_-prefixed IDs (148,527 rows) come from the newer run. The two are identical in format and were validated identically — the only difference is the seniority mix, which is flatter in the older batch (0.3% staff and executive, versus 5.9% and 2.1% in the newer one). Filter on the prefix if you need the more balanced subset.

Intended use

Training and evaluating candidate–job relevance models — cross-encoder rerankers, bi-encoder retrieval, matching heuristics. Useful anywhere real candidate data is unavailable or shouldn't be used.