sky9262/laya-jd-resume-ranking-enja

Laya Bilingual JD–Résumé Ranking (EN/JA)

1,200 (job description, resume) pairs in English and Japanese, each labelled by a three-model ensemble with soft probability distributions rather than hard classes. Built to fine-tune Laya — Convai's 421M-parameter ModernBERT decision head — so it scores a pool of candidates against one job description and ranks them.

The dataset is fully synthetic. No real résumé, no real person, no real employer relationship is represented.

This release replaces the earlier 24-group / 720-case version. The corpus now carries 40 JD groups and 1,200 labelled cases, and the splits are language-balanced as well as group-wise. Every number below is measured on the 1,200-case corpus.

Warning — no model trained on this dataset has been shown to work on real résumés. A fine-tune of this corpus scores well on this corpus's own held-out splits and then fails on real data: on 100 real Japanese résumés, 63 candidates holding the exact target role ranked at a median position of 49 of 100 — chance level — and a control job description that nobody in the pool fits scored its top candidate 2.99, higher than any candidate scored against the genuinely matching JD (2.32). The full figures and the leading explanation are in the Reference results section below. Treat this corpus as research material for ranking and distillation, not as training data for applicant screening.

What this trains

Target model	convaiinnovations/laya — 421M, ModernBERT-large decision head
Fine-tune measured below	an unreleased ModernBERT decision-head fine-tune of this corpus, called v5 throughout this card — see Reference results, including its failure on real résumés
Multilingual checkpoint	convaiinnovations/laya subfolder multilingual — mmBERT-base, 322M, 100+ languages
Apple Silicon inference	aac6fef/laya-mlx, aac6fef/laya-multilingual-mlx
Reference fine-tune	typed-decisions notebook
Benchmark this imitates	LocalLLaMA/typed-decisions

Laya is a non-autoregressive typed-decision model: one bidirectional forward pass returns probabilities over options you supply, with output_tokens: 0. It answers choice, score and noul (boolean) questions and cannot hallucinate an option that is not in your list.

Base Laya scores 0.362 zero-shot on the typed-decisions benchmark; the published fine-tune reaches 0.766. This dataset is the equivalent training corpus for JD-conditioned résumé ranking, a task base Laya was never trained for.

The record format matches what Laya's trainer consumes directly — state, questions, gold — so no conversion is needed.

The task

Given one JD and 30 candidate résumés, score every candidate and order them. Three typed questions per pair:

question	type	options
match	score	0 no match · 1 weak · 2 partial · 3 good · 4 excellent
skills_match	score	0 none · 1 minimal · 2 some overlap · 3 strong overlap · 4 near exact
seniority_fit	choice	underqualified · good_fit · overqualified · wrong_field

match is the ranking signal. wrong_field exists because a senior person in an unrelated profession is simultaneously over-levelled and under-qualified, and a three-way choice cannot express that — see Construction.

Shape

1200 cases = 40 JD groups × 30 candidates
400 bilingual résumés (EN + JA for the same person), 43 occupations
258 job descriptions in the bank, 40 used
3600 teacher votes (3 models × 1200 — 1200 each from opus, sonnet, fable)

The 400 résumés are unchanged from the 720-case release; they are reused across more JDs rather than regenerated. Median 3 appearances per résumé, max 5. Each appearance judges the same person against a genuinely different role.

Splits are by JD group, so no job description appears in two splits and the test roles are genuinely unseen. The disjointness is asserted in code. Splitting by case would leak: the same JD's other 29 candidates would sit in training.

split	cases	groups	en_en	en_jd_ja_resume	ja_ja	ja_jd_en_resume
train	690	23	251	139	231	69
validation	270	9	96	54	93	27
test	240	8	77	43	93	27

Splits are also language-balanced, not only size-balanced: ja_ja is 93 cases in both validation and test, and ja_jd_en_resume is 27 in both. This matters more than it looks. A size-only split put 48% ja_ja in test against 29% in validation — which would have manufactured a validation/test performance gap that had nothing to do with the model, only with which language mix each split happened to draw. Balancing the language composition makes the two evaluation sets comparable to each other.

Record format

{
  "state": "### JOB DESCRIPTION\n…\n\n### CANDIDATE RESUME\n…",
  "questions": {"match": {"type": "score", "instructions": "…", "criteria": [...]}, "…": {}},
  "gold": {"match": {"probabilities": {"0": 0.33, "1": 0.67, "2": 0.0, "3": 0.0, "4": 0.0}}, "…": {}},
  "case_id": "case-00001", "jd_id": "jd-…", "lang_pair": "en_jd_ja_resume",
  "jd_variant": "brief", "pairing": "partial"
}

gold.probabilities is the vote spread across three teachers, not one model's opinion. A 2–1 split becomes {2: 0.667, 3: 0.333}. Train against the distribution (KL / soft cross-entropy), not the argmax.

State length: min 461, median 683, max 999 tokens. A 1024-token context truncates nothing.

Coverage

• Languages — en_en 424, ja_ja 417, en_jd_ja_resume 236, ja_jd_en_resume 123. Cross-lingual pairs are 30% of the corpus: an English posting against a Japanese 職務経歴書 is a real scenario in Japanese hiring and the hardest case for a ranker.
• JD length — title 450, full 390, brief 360. A JD may be a full posting, three lines, or a bare job title; all three must work.
• Relevance — strong 244, partial 596, weak 360. Deliberately oversampled toward contested pairs (see Known limitations).

Japanese résumés are 職務経歴書 with conventional sections (職務要約 / 活かせる経験・知識・技術 / 職務経歴 / 資格 / 自己PR) across three formats — 逆編年式 204, 編年式 105, キャリア式 91. Career shapes vary: stable 202, job-hopper 91, career-changer 70, employment gap 37.

Label quality

Overall teacher agreement 0.895 (mean modal vote share across 3 models). The original 720-case release measured 0.887 on the same statistic, so the second labelling round is consistent with the first — no drift between rounds. All 1,200 cases carry votes from all three teachers, and the panel is exactly balanced at 1,200 votes each from opus, sonnet and fable.

question	modal-vote share (unweighted)	unanimous
match	0.883	66.0%
skills_match	0.880	64.8%
seniority_fit	0.922	77.8%

34.0% of cases have at least one teacher disagreeing on match. Only 9 argmax_tie flags across all 1,200 cases — the aggregate distribution has a clear modal answer nearly everywhere.

slice	match agreement
weak pairs	0.965
strong pairs	0.851
partial pairs	0.847
title JD	0.901
full JD	0.881
brief JD	0.862

Two results worth reading carefully. partial is the most contested bucket, not strong — adjacent-occupation candidates generate more genuine disagreement than either clear matches or clear misses. And brief JDs are harder than bare titles: a title forces teachers to reason from the role itself, while three ambiguous lines invite divergent readings.

gold.match distribution: 0:630 · 1:232 · 2:142 · 3:91 · 4:105 — 47.5% non-zero. Per group, median 5 candidates at gold ≥ 3 (min 0, max 11); 38 of 40 groups have genuinely good candidates to surface.

gold.seniority_fit: wrong_field 810 · good_fit 173 · overqualified 162 · underqualified 55. Overqualified outnumbers underqualified roughly 3:1 — for a ranker, over-levelled candidates are the dominant failure mode.

Reference results — in-distribution only, and they did not transfer

Read this section before training on this corpus. The numbers further down are real and correctly measured on this dataset's own held-out splits. They did not survive contact with real résumés.

What happened on real résumés

v5 — a ModernBERT decision-head fine-tune of this corpus — was evaluated on 100 real Japanese résumés:

• 63 candidates holding the exact target role ranked at a median position of 49 of 100. That is chance. Holding the advertised job title moved a candidate no higher than a coin flip would.
• A control JD that nobody in the pool fits scored its top candidate 2.99 — higher than any candidate scored against the genuinely matching JD, whose best was 2.32. The model scored a job description it should have rejected outright above the one it should have matched.
• Reformatting the résumé input as prose did not help: median rank 54 against 49. The failure is not an input-formatting artifact.

The leading explanation

In this corpus a strong pairing was constructed by matching a résumé's seed_category to the JD's jd_category, and that résumé's text was generated from that same category. Category membership is therefore written into the résumé wording itself, so a model trained here can reach high in-distribution scores by learning generator-injected vocabulary rather than role semantics. The teacher ensemble then labelled that same synthetic text, so the labels certified the artifact instead of correcting it. Real résumés carry no such injected vocabulary, and the learned signal has nothing to fire on.

The in-distribution numbers

v5 pooled over 510 cases / 17 JD groups of this dataset's own held-out splits:

metric	value
median Spearman	0.6094 (std 0.2010)
NDCG@5	0.7716
NDCG@10	0.7840
NDCG@20	0.8464
precision@1 (mean)	0.5714
match accuracy	0.5176

Ranking quality runs well ahead of pointwise accuracy, which is the expected shape for a model trained on soft targets: the ordering is learned better than the absolute grade. Read these as a measure of fit to this corpus. Given the real-résumé results above, they are not evidence of screening ability, and a higher score on this table is not evidence that the next model will transfer.

Construction

1. Seeds — 1200 structured content seeds (titles, skills, education, career spans) extracted from ahmedheakl/resume-atlas (MIT). That corpus is lowercased, punctuation-stripped and stopword-stripped — a bag-of-words stream, unusable as résumé text — so it was mined for content, not prose.
2. Generation — an LLM wrote each candidate's facts once (name, employers, dates, seniority), then authored the English résumé and the Japanese 職務経歴書 independently from that fixed brief. Facts cannot drift between languages by construction. All identities are invented, which also removes the source corpus's PII.
3. JD bank — 258 postings, Japanese written as Japanese postings (【職種】【仕事内容】【応募資格（必須）】…, です・ます調) rather than translated.
4. Pairing — 40 JDs, each with 30 candidates mixing strong / partial / weak relevance and both résumé languages, stratified within each relevance bucket so language is not confounded with difficulty.
5. Labelling — three models (Claude Opus, Sonnet and Fable) scored every pair blind: each teacher saw only case_id, jd_text, resume_text. The pairing label, relevance prior and category fields were stripped from the input files rather than the teachers being asked to ignore them.
6. Aggregation — vote spread → probability distribution, with per-question agreement and total-variation recorded.

Hard-case families

The teachers surfaced these unprompted; they are the cases a keyword matcher fails.

• Title collisions — an attorney in civil litigation against a Civil Engineer post; an enterprise architect against a building-architecture role; a Food and Beverage Distribution Supervisor (truck logistics) against an F&B service role; English advocate (patient/child advocate) against Japanese 弁護士 (licensed attorney).
• Trade vs engineering — a 32-year Master Electrician against a circuit-design engineering post. Right domain vocabulary, different profession.
• Adjacent stack — a lead SQL Server DBA against a .NET engineer role: owns one requirement outright, writes no C#.
• Stale credentials — genuine practitioners whose last relevant role ended a decade ago.
• Overqualification — in the junior-role groups, no candidate earns a 4, because every strong profile is a manager applying to an entry seat. 8 of the 40 groups contain no grade-4 candidate at all.

Using it

from datasets import load_dataset
d = load_dataset("sky9262/laya-jd-resume-ranking-enja")

Train against the distribution, not the argmax — gold[q]["probabilities"] is the teacher vote spread and the disagreement is the signal. Laya's own trainer consumes these records unchanged.

At inference

state = f"### JOB DESCRIPTION\n{jd_text}\n\n### CANDIDATE RESUME\n{resume_text}"
out   = agent.predict(state, questions)

out["answers"]["match"]["score"]          # 3.6693 — continuous, this is what you sort by
out["answers"]["match"]["probabilities"]  # {"0":0.003, "1":0.029, ... "4":0.804}
out["answers"]["seniority_fit"]["choice"] # "good_fit"
out["usage"]["output_tokens"]             # 0 — nothing is generated

The state format is unchanged from the previous release.

Ranking a pool is the loop:

scored = [(r, agent.predict(build_state(jd, r), questions)) for r in resumes]
ranked = sorted(scored, key=lambda x: -x[1]["answers"]["match"]["score"])

score returns a continuous interpolated value rather than a class index, which is what makes it usable as a sort key. confidence is a separate quantity from max(probabilities) and the two diverge — use the distribution to rank, treat confidence as a review flag.

Keep the framing consistent

The ### JOB DESCRIPTION … ### CANDIDATE RESUME wrapper is what the model is calibrated on. Raw text still works, but the answer moves. Measured on one case, six framings of identical content:

framing	match	seniority_fit
trained format	3.291	good_fit
bare concatenation	3.357	good_fit
Job: / Candidate: labels	3.320	overqualified
【求人】/【応募者】 labels	3.178	overqualified
résumé first, JD second	2.715	overqualified
dict {job_description, resume}	3.030	overqualified

A 0.64-point spread on match and a flipped seniority_fit. Reversing the order costs most — position is how the model tells requirement from applicant.

Ranking compares candidates against the same JD, so a constant offset cancels out. What breaks a ranking is framing candidates differently from each other — which is exactly what happens when some résumés arrive as clean text and others from a PDF that extracted with different spacing. Normalise on the way in.

PDFs

Laya takes text, so extract first. pdftotext -layout handles Japanese 職務経歴書 cleanly — section headers survive, no mojibake, and -layout preserves the date/employer columns résumés use.

text = subprocess.run(["pdftotext", "-layout", path, "-"],
                      capture_output=True, text=True).stdout

That covers text-based PDFs. Scanned or photographed résumés need OCR (tesseract -l jpn), as do 履歴書 submitted as filled-in image templates, which are common in Japan. Guard for it: extraction returning under ~200 characters almost always means a scan, and should route to OCR rather than through the model as near-empty text.

Extending it

If you need more data, add groups, not résumés. The résumés are the expensive asset and this corpus still does not exhaust them. This release is the 3× step described in the previous card, carried out.

JD bank	258 written, 40 used — 218 sit unused
Résumé reuse	3.00 groups per résumé (median 3, max 5)
this release, 3× reuse	40 groups, 1,200 cases
at 4× reuse	53 groups, 1,590 cases
at 5× reuse	66 groups, 1,980 cases

Expanding groups costs labelling only — no generation. Each additional appearance judges the same résumé against a genuinely different role, so 3–4× reuse is normal for a ranking corpus.

Going from 24 to 40 groups fixed the weakest part of the previous release: the test split was 4 groups, a median Spearman over four numbers where one flat group cost a quarter of the evaluation. It is now 8 groups, and training carries 2,070 decisions (690 cases × 3 questions) against the 1,440 of the previous release.

When you expand further, draw the new groups from the occupations with the most résumés. Depth is what produces contested pairs; spreading thinly across all 43 occupations is what produced the flat groups here.

Known limitations

Read these before trusting a number from this dataset.

• The grade distribution is bottom-heavy. Roughly half of every 30-candidate pool is gold grade 0 (mean 15.8 of 30), and only about 3 are grade 4 (mean 2.6). The interesting part of each ranking is a small slice of the pool.
• Soft targets are narrower than ideal for distillation. Teachers agree within ±1 grade about 99% of the time (98.9% on match), so the vote spread carries less information than a genuinely divided panel would. And agreement measures consistency, not correctness — three models trained on overlapping data can be consistently wrong together. No human has validated these labels.
• Agreement is 0.895, higher than intended. The reference corpus this imitates sits near 0.646. Teachers agree here because most cross-occupation pairs genuinely are zero, not because the questions are trivial — but the corpus is easier than planned. Excluding weak pairs it drops to roughly 0.85.
• Full-length JDs are the hardest variant for models trained on this data. v5 scores 0.450 on full against 0.734 on brief and 0.691 on title. More JD text makes ranking worse, and it is not truncation — the longest input in the corpus is 999 tokens against a 1024 window. The cause is not established.
• Occupational depth is thin. 400 résumés over 43 occupations is ~9 per occupation, so some JDs have few genuine matches. 2 of 40 groups contain no candidate above gold 2, and per-group NDCG will be uneven.
• Relevance mix is not a production base rate, and it missed its own target. strong/partial/weak was specified at 30/40/30 to oversample contested cases. The pairing stage fell 45 strong slots short on the 16 new groups — realised strong share 99/480 ≈ 21%, and 244/1200 = 20.3% corpus-wide against a 30% target. A real applicant pool is in any case mostly non-matches. Absolute scores may need recalibration on a production-rate holdout; ranking is invariant to monotone recalibration, which is why ranking is the headline metric.
• Résumés are LLM-generated, not real. Real résumés — PDFs, tables, inconsistent formatting, idiosyncratic section ordering, typos — are unrepresented here. A model that ranks this corpus well has not been shown to rank real submissions well.
• Source category labels are noisy. Some seeds are mislabelled in the source corpus (a "Data Science" seed whose titles are all K-12 teaching). Generation followed the actual career, so pairing — derived from the category string — carries noise. It is a stratification prior, not a label.
• Japanese is not native-reviewed. Written and self-checked by language models against conventional structure, numerals (万/億) and katakana. No native speaker has audited it.
• Synthetic throughout. Résumé prose, JDs and labels are all model-generated. Distributional quirks of the generating models are present.

Intended use and misuse

Built for research on ranking and typed decision-making over bilingual documents, and as distillation data for small encoder-based decision models.

Not suitable for deployment in real hiring decisions. It is synthetic, its labels are model opinions rather than hiring outcomes, and it has had no fairness audit. Automated employment screening is regulated in several jurisdictions — the EU AI Act treats it as high-risk, NYC Local Law 144 requires bias audits. A model trained on this data should be treated as a research artifact.

One deliberate property: candidate names are uniform in style across the corpus, so there is no name-derived signal for a model to learn. That removes one discrimination vector; it does not make a model trained here fair.

Privacy and contact details

All résumés are synthetic. No contact detail in this dataset routes anywhere real.

Email — every address uses an RFC 2606 reserved documentation domain: example.com (811 occurrences) and example.jp (141). Zero addresses on any real provider.

Telephone — this is a correction to the previous release. This version normalises all fictional numbers into the NANP reserved range 555-0100 through 555-0199, which is the only block guaranteed never to be assigned. The previous release contained 20 distinct numbers in 555-02xx and above. Those are technically assignable and therefore not guaranteed fictional. 24 distinct numbers were remapped across 26 text fields in 400 résumés. Verified after remapping: zero occurrences outside the reserved block.

One honest consequence: the remapping happened after v5 was trained, so the résumé text published here differs from what v5 saw — by those phone digits and nothing else. It does not affect any label, any split assignment, or any metric reported above.

Provenance

Content seeds derive from ahmedheakl/resume-atlas (MIT). All prose, all job descriptions and all labels in this dataset are newly generated. Released CC-BY-4.0.