anatiagent/resume-posting-pairs

Résumé–posting pairs

Résumé–posting pairs labelled for fit — high, medium or low — by Jev (typesafe/jev-1.13, questions q1), from an MSc project on résumé screening. Each résumé is in one pair only. Two sets, loaded as separate configs so they never mix:

config	file	pairs	high / medium / low	split
main (default)	pair_dataset_resume_posting1110.csv	16,440	5,480 each	train
heldout_scraped_2026-10	heldout_scraped/2026-10/pair_dataset_resume_first_jev-1.13_q1_cap10_minp0.6_rw50_pw30_unbalanced.csv	747	154 / 293 / 300	test

from datasets import load_dataset

main = load_dataset("anatiagent/resume-posting-pairs", "main", split="train")
heldout = load_dataset("anatiagent/resume-posting-pairs", "heldout_scraped_2026-10", split="test")

The two sets

main. Résumés from the Florex résumé corpus (Indeed, 2019) against IT postings from LinkedIn (2024) and Dice.com (2017), in seven domains. The three tiers are the same size with the same domain mix, and about 3% of the pairs are off-domain lows (a résumé against a posting from an unrelated domain).

heldout_scraped_2026-10. A held-out test set from data scraped in 2026: résumés people posted to Reddit, read by OCR, against tech postings from Israeli job boards. It went through the same extraction, domains, candidate draw and judge as main, with these differences:

• eight domains — six of main's seven (no Networking) plus Quality Assurance and Data Science;
• unequal tiers — each résumé once, the scarcest tier filled first, so the tiers differ in size and in domain mix (Project manager has no high pairs);
• no off-domain pairs — every pair is a résumé against a posting of its own domain;
• résumé domains come from a model rather than from curated labels.

More drops may be added later, each as its own config.

Columns

columns	meaning
pair_id, job_id, resume_id	ids; strings
job_domain, resume_domain, pair_type, pool	the pair's domains, same or off domain, and the pool it was selected from
fit	the label: high, medium or low
jev_p_high, jev_p_medium, jev_p_low, jev_p_fit, jev_margin	Jev's probability on each level, on the chosen one, and its lead over the second
jev_share	Jev's estimate of the share of the posting's must-haves the résumé shows
model, questions_version	the judge, exactly
band, rank, cosine, skill_coverage, skills_required	how the candidate was drawn: embedding similarity (text-embedding-3-large) and skill overlap — for baselines, not labels
years_*, degree_*	years of experience and degree level, required and stated; empty where unknown
job_title, job_qualifications, resume_education, resume_qualifications	the four texts Jev read; the qualifications are LLM extractions (openai/gpt-oss-20b)
job_description, resume_text_clean	the full originals they were extracted from

Empty cells are empty strings, not nulls; numeric columns that can be empty are typed as strings.

Labels and limits

The labels are an LLM judge's, not a recruiter's. Pairs where Jev's probability on its own level was under 0.6, and pairs with a résumé under 50 words or a posting under 30 words of qualifications, were left out.

Licence

other: each source dataset keeps its own terms — see the links above for main; the heldout_scraped data was scraped from public pages. Use for research.