README.md

metadata

license: mit
language:
  - en
size_categories:
  - 10K<n<100K
task_categories:
  - text-ranking
tags:
  - synthetic
  - jobs
  - recruiting
  - resumes
  - learning-to-rank
pretty_name: Synthetic US Candidate Profiles
configs:
  - config_name: default
    data_files: profiles.parquet

Synthetic US Candidate Profiles

99,998 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.

Quick start

from datasets import load_dataset
import json

ds = load_dataset("<your-org>/<your-dataset>", split="train")
profile = json.loads(ds[0]["profile_json"])
print(profile["identity"], profile["skills"]["primary"])

Schema

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

column	type	description
profile_id	string	unique identifier
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
profile_json	string	complete profile as a JSON object

profile_json contains:

key	contents
identity	seniority, job_family, years_experience, is_student, is_career_changer
location	city, state, country, open_to_remote, open_to_relocation, preferred_locations
objective	career_interests, desired/min salary, employment_types, target_company_tags
eligibility	work_authorization_countries, sponsorship_needed, legal_status, security_clearance
skills	primary, secondary, languages
experiences	title, company, industry, size, start/end dates, is_current, is_remote
educations	school, degree, field, dates
certifications, projects	list records, often empty
source_meta	llm_model, cohort, seed_role, which fields were sampled rather than generated

years_experience is computed from date spans, never asked of the model — LLMs are unreliable at date arithmetic. eligibility and open_to_remote are sampled from a fixed distribution rather than inferred, both for controlled coverage and because guessing someone's visa status is not something a model should be doing.

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	19,998		manager_lead	7,000
entry_level	13,000		over_qualified	6,000
career_changer	10,000		return_to_work	5,000
sparse_profile	9,000		visa_constrained	5,000
senior_specialist	9,000		bootcamp_self_taught	3,000
student_intern	8,000		military_transition	3,000
			gig_contract	2,000

Job families (14) — other 19.0%, operations 17.7%, software_engineering 15.8%, healthcare 10.3%, research 6.3%, hardware_engineering 6.2%, finance 4.9%, education 3.8%, design 3.7%, marketing 3.3%, legal 3.2%, sales 2.6%, and 2 more.

Seniority — senior 31.5%, entry 27.2%, mid 24.8%, junior 5.0%, manager 4.3%, intern 3.2%, principal 2.1%, director 1.4%, staff/executive 0.5%.

Content — 803 distinct seed occupations, 66 states/territories (top: California, Texas, Florida, Colorado, North Carolina, Arizona). Median 13 skills, 3 experience entries, 10.2 years experience (p10 1.9, p90 20.2). 69.8% open to remote, 5.1% need sponsorship, median target salary $85,000.

How it was built

Roles were enumerated systematically rather than sampled, so coverage is even rather than popularity-weighted: a vocabulary bank was generated per occupation (803 roles → 36,040 distinct skill, tool, specialization, and certification terms), then each profile was seeded with a role, cohort, city, and skill subset before generation at temperature 1.0. Every profile was validated against a JSON schema; 14 were rejected and 2 more dropped in verification.

Generation cost roughly $24 and 140 minutes at 96-way concurrency.

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.

Limitations

• Synthetic. Reflects an LLM's model of what job seekers look like, not the real distribution. Validate on real data before trusting production numbers.
• US only. Every profile is US-based. Do not expect transfer to other labour markets.
• Round-number bias. LLM-generated salaries and dates cluster on round values.
• Inherited bias. Name, location, industry, and seniority associations reflect the generating model's biases. Not suitable for fairness auditing, and any hiring-adjacent use needs its own bias evaluation.
• state is a full name ("Illinois", not "IL"). Normalize before joining against systems that store two-letter codes.
• Not a hiring tool. These are training fixtures, not candidates.