HeadcountLabs/job-postings

Headcount Labs Job Postings

Job postings collected directly from employers' own hiring systems: applicant tracking systems (ATS), company careers pages and public government job exchanges. Each posting keeps the structured fields the employer published and a link to the original listing. Job description text is not included in this release. Postings of the same opening are linked to one stable job_id.

Source: Headcount Labs (headcountlabs.com), licensed under CC BY 4.0.

This is a snapshot of the data served by the public Headcount Labs JobCensus API. Snapshot taken 2026-10-14 5:28 PM ET to 2026-10-14 8:05 PM ET, change sequence 22659803.

Table	What it holds	Rows
jobs	one row per job	4,871,084
postings	one row per posting (a listing of a job on one source)	5,716,621
job_locations	one row per job location, with country, region, locality, zip and metro area code	5,583,156
companies	one row per company ID seen on a job (a summary, not a company register)	12,073

Postings date from 2005-08-08 to 2026-12-09 (posted_at). The data is not a sample: it is every job the API serves at the snapshot time.

What the data is

• Source. We read the public job boards of employers over more than 100 ATS platforms, company careers pages and government job exchanges. Only postings whose source terms allow redistribution are included, and a job needs at least one such posting.
• Employer-written text. Job titles were written by the employers who posted the jobs, not by us. We do not edit them. Pay, location and dates are copied as published, or derived by us as marked in SCHEMA.md.
• Source or derived. Every descriptive value carries a state that says whether the source published it (known, source_omitted, source_unmapped) or we computed it (derived). Nothing a source published is dropped.
• No description text. This release leaves out the job description text. has_description says whether the source published one, and posting_url links to the original listing, where the employer's full text is.

Files

Folder	Files	Size	Rows
data/companies/	1	380.0 KB	12,073
data/job_locations/	1	159.3 MB	5,583,156
data/jobs/	228	1.5 GB	4,871,084
data/postings/	219	3.2 GB	5,716,621

Total 4.9 GB. Parquet is zstd compressed with row groups of about 100,000 rows and files under 1 GB. Files in data/jobs and data/postings are named part-YYYYMM-N.parquet and hold the rows whose posted_at falls in that month. For jobs, posted_at is the earliest posting date of the job. For postings it is the source's posted date when it is a plausible date (2000 to 2027), otherwise the date we first saw the posting (posted_at_basis says which).

Columns per table: jobs 21, postings 34, job_locations 7, companies 3. The columns, types, meaning and origin are in SCHEMA.md. Several columns (*_json) hold a JSON object with more parts inside, described there. MANIFEST.json lists every file with its row count, byte size and SHA-256, and the snapshot details.

Coverage

• Snapshot of the whole public job set at 2026-10-14 8:05 PM ET. The export reads a live database, so jobs that changed during the export were read again afterwards. 1,105,961 change-log entries were written while it ran (see MANIFEST.json, snapshot).
• Most postings are from the United States. country in job_locations and the location fields give the country for each job.
• Languages. Most postings are in English. Language share, measured on a sample of job descriptions before they were left out of this release:

Language (ISO 639-1)	Share of sampled descriptions
en	83.8%
de	3.7%
fr	3.5%
pt	2.6%
zh-cn	1.8%
es	1.1%
ja	1.1%
nl	0.8%
other	1.5%

Sample: 4,941 random job descriptions with more than 100 characters, detected with langdetect.

• Fill by source type. Share of postings whose source published a description, a pay range (pay_min or pay_max), a location (location_json state known) and a company link (the job has a company_id), for the 25 largest source types:

Source type	Postings	Source has description	Pay	Location	Company link
workday	750,297	99.8%	36.0%	100.0%	30.9%
oracle_hcm	392,932	99.9%	52.0%	100.0%	26.0%
successfactors	333,933	99.9%	11.3%	100.0%	17.1%
icims	292,624	100.0%	44.6%	100.0%	17.0%
adp	283,784	99.9%	33.7%	100.0%	10.7%
smartrecruiters	212,860	100.0%	11.6%	100.0%	16.2%
ukg	197,884	100.0%	28.8%	100.0%	5.4%
jibe	192,038	100.0%	34.7%	100.0%	33.2%
greenhouse	190,962	99.9%	48.9%	100.0%	8.9%
taleo	189,541	99.8%	17.0%	100.0%	27.1%
phenom	175,958	99.9%	32.9%	100.0%	40.3%
directemployers	163,515	100.0%	58.5%	100.0%	23.0%
paycom	122,055	99.9%	57.1%	100.0%	7.2%
dayforce	115,171	99.8%	17.7%	100.0%	29.7%
radancy	98,630	100.0%	35.3%	100.0%	45.0%
paylocity	98,520	100.0%	37.7%	100.0%	4.0%
(none)	92,403	98.2%	39.9%	100.0%	33.5%
beisen	82,398	100.0%	0.2%	100.0%	0.0%
jazzhr	80,954	100.0%	44.4%	100.0%	1.3%
careerplug	76,494	100.0%	76.7%	100.0%	5.2%
csod	75,998	99.9%	17.7%	100.0%	8.3%
gupy	69,680	100.0%	0.0%	100.0%	0.0%
zoho_recruit	69,636	99.8%	6.4%	100.0%	0.0%
employer_site	67,999	99.0%	49.1%	100.0%	11.6%
ashby	63,024	100.0%	37.9%	100.0%	20.1%
all	5,716,621	99.4%	33.5%	100.0%	16.9%

The full list is vendor_fill.csv.

Update plan

The dataset is a snapshot. We plan to add daily change files under data/changes/date=YYYY-MM-DD/ and a full refresh every week, keeping the file layout, so you can sync only what changed. The snapshot block in MANIFEST.json tells you which version you hold. Live data is also available through the Headcount Labs JobCensus REST API and MCP connector.

How to load it

Hugging Face datasets (streaming, no full download):

from datasets import load_dataset

jobs = load_dataset("HeadcountLabs/job-postings", "jobs", split="train", streaming=True)
for row in jobs.take(3):
    print(row["title"], "|", row["employer_name"], "|", row["posted_at"])

pandas (one month):

import pandas as pd

df = pd.read_parquet("data/jobs/", filters=[("posted_at", ">=", "2026-10-01")])
print(df[["title", "employer_name", "work_arrangement"]].head())

DuckDB (local files, or hf://datasets/HeadcountLabs/job-postings/...):

SELECT employer_name, count(*) AS jobs
FROM read_parquet('data/jobs/*.parquet')
WHERE employer_name IS NOT NULL
GROUP BY 1 ORDER BY 2 DESC LIMIT 10;

-- jobs per state
SELECT region, count(DISTINCT job_id) AS jobs
FROM read_parquet('data/job_locations/*.parquet')
WHERE country = 'US' GROUP BY 1 ORDER BY 2 DESC LIMIT 10;

-- share of postings that publish pay, by month
SELECT substr(posted_at, 1, 7) AS month, count(*) AS postings,
       round(avg((pay_min IS NOT NULL OR pay_max IS NOT NULL)::int), 3) AS share_with_pay
FROM read_parquet('data/postings/*.parquet')
GROUP BY 1 ORDER BY 1;

Amazon Athena (replace BUCKET):

CREATE EXTERNAL TABLE jobs (
  `job_id` string,
  `membership` string,
  `merged_into_job_id` string,
  `currentness` string,
  `company_id` string,
  `employer_name_json` string,
  `title_json` string,
  `work_arrangement_json` string,
  `locations_json` string,
  `employer_name_key` string,
  `work_arrangement` string,
  `job_first_seen_at` string,
  `curated_at` string,
  `job_updated_at` string,
  `last_change_seq` bigint,
  `has_description` bigint,
  `has_location` bigint,
  `title` string,
  `employer_name` string,
  `posted_at` string,
  `posting_url` string
)
STORED AS PARQUET
LOCATION 's3://BUCKET/data/jobs/';

CREATE EXTERNAL TABLE postings (
  `job_id` string,
  `posting_id` string,
  `board_id` string,
  `source_posting_id` string,
  `posting_status` string,
  `missing_from_latest_fetch` bigint,
  `posting_url_json` string,
  `application_url_json` string,
  `location_json` string,
  `pay_json` string,
  `title_json` string,
  `employer_name_json` string,
  `work_arrangement_json` string,
  `source_posted_at` string,
  `first_seen_at` string,
  `last_completed_fetch_at` string,
  `latest_application_check_json` string,
  `semantic_hash` string,
  `source_data_json` string,
  `source_rights_json` string,
  `internal_source` bigint,
  `board_id_filter` string,
  `source_type` string,
  `employment_type` string,
  `source_class` string,
  `board_key` string,
  `posting_url` string,
  `application_url` string,
  `pay_min` double,
  `pay_max` double,
  `pay_currency` string,
  `pay_interval` string,
  `posted_at` string,
  `posted_at_basis` string
)
STORED AS PARQUET
LOCATION 's3://BUCKET/data/postings/';

License

CC BY 4.0: share and adapt, including commercially, with attribution.

Source: Headcount Labs (headcountlabs.com), licensed under CC BY 4.0.

Job posting text belongs to the employers that published it. Where a posting's attribution field says credit is required, show that credit with the job. The field is source_rights_json (attribution_required, attribution_text, attribution_link).

Licence scope

CC BY 4.0 covers Headcount Labs' compilation and the fields we derive: normalised titles, locations, pay, dates, company links and deduplication. Job description text, employer names and employer logos stay with their owners. This release does not include description text. Employer names are included to identify each posting, and every posting links to the original listing (posting_url, application_url).

Takedown

Employers or individuals can ask for a posting or an employer to be removed by writing to support@headcountlabs.com.

Questions and corrections: headcountlabs.com.

Known limits

• Fields are as published by each source. Pay can be hourly, yearly or in any currency (see pay_interval, pay_currency); it is not converted.
• currentness describes our evidence about whether a job is still open at the snapshot, not a guarantee.
• The stale flag is not in this version. On the snapshot date it was wrong for a large share of jobs (a 30-day first-listed rule, since replaced by source freshness). It returns in the next version.
• The export reads a live database over several hours. In a random check of 200 jobs against the source, 188 (94%) matched in every column. The rest had been updated at the source after they were read (refreshed fetch times, employer links, location or title text).
• Email addresses and phone numbers were removed from job titles, employer names and location text (street lines, text_key) in this release. application_url values that are mailto: links are kept, because they are the way the employer asks people to apply. Phone-like digit runs inside URLs and source ids are left as published.
• A few posted_at values are in the future (the source published a date after the snapshot). They are kept as published.
• A small number of text values with invalid UTF-8 in the source were stored with replacement characters (count in MANIFEST.json).