JobOpportunitiesAPI/joa-extraction-arena

JOA Job-Posting Extraction Arena

Disclosure: Job Opportunities API (JOA, jobopportunitiesapi.org) is an independent data business that sells API access to job-posting data. AI helped run the experiments, check the numbers and draft this text; Loukas (Luca) Tzekos is editorially responsible. Contact: hello@jobopportunitiesapi.org.

A benchmark for one narrow task: reading a real job posting and filling 11 structured fields. Job Opportunities API (JOA) built it to choose and train its own extraction models. This repository holds the results for 84 models (68 open-weight models, 13 quantised variants and JOA's 3 fine-tuned models), the per-field scores, the grading method and the reference labels for the listings we are allowed to share (221 of the 250).

Article: https://dev.to/lucagiftzek/teaching-small-models-to-read-job-postings-our-arena-and-the-joa-small-advance-and-deep-fine-tunes-3hpb

The task

Each model receives the text of one job posting, as stored by JOA, plus any structured data found on the page, and must return 11 fields:

Field	Meaning
title	The posting title as published
company	The hiring organisation as named in the posting
location	Where the job is, as stated
salary_currency	ISO currency code of the stated pay
salary_min	Lower bound of the stated pay (equal to salary_max when one figure is given)
salary_max	Upper bound of the stated pay
salary_period	hour, day, week, month or year
seniority	Entry Level (0 years), Entry Level (1 year), Mid (2-3 years), Senior (4+ years), Lead, Manager, Director or Executive
remote_type	onsite, hybrid or remote
posted_date	Date the posting was published (ISO 8601)
is_agency	true when the poster is a staffing agency or third party rather than the employer

A field the posting does not state should be left empty. Empty is a valid answer.

How it was run

• Listings: 250 real, open job postings sampled from JOA in September 2026, from employers' own recruiting systems and careers pages.
• Same input for everyone: every model received the same prefetched page text and structured data, with a completion budget of up to 24,000 tokens. A reply cut off mid-thought counts as no answer.
• Reference labels: built in September 2026 with Claude Opus agents that read JOA's stored text and the live page, with a confidence level per field. Each cell also records whether its value could be found in the input the models were shown.
• Two judges: each answer was graded field by field by two open models, Kimi-Linear-48B-A3B-Instruct and Seed-OSS-36B-Instruct, against the reference. A cell's score is the mean of the two grades. Both judges also appear as contestants in the table.
• Scoring rules: rule A (the main score) leaves out cells whose reference value appears nowhere in the model's input and that almost no model got right; a missing answer where the value could be found scores 0. Rule B (strict) scores every missing answer as 0. Rule C (lenient) leaves out every "no answer, uncertain reference" cell. In the run of 25 September 2026, rankings under the three rules correlated at 0.99.
• Uncertainty: 95% ranges come from a paired bootstrap over listings (1,000 resamples). Re-running the same model at temperature 0 moved its score by up to 1.5 points, so treat smaller differences as noise.
• Dates: stock and quantised models graded 23–29 September 2026, JOA Advance 29 September 2026, JOA Small and JOA Deep 2 October 2026. None of the 250 listings is in the fine-tunes' training data.

Top of the scoreboard (rule A, 84 models, as of 2 October 2026)

Rank	Model	Score	95% range	Answered	Median output tokens
1	JOA Small (fine-tuned ibm-granite/granite-4.2-3b)	93.3	92.2–94.4	100%	117
2	JOA Advance (fine-tuned ibm-granite/granite-3.3-8b-instruct)	92.7	91.6–93.9	100%	123
3	JOA Deep (fine-tuned Qwen/Qwen3.8-27B)	92.2	91.0–93.4	100%	111
4	Qwen/Qwen3.5-27B (the fine-tunes' trainer)	89.2	86.9–91.4	98%	2,946
5	Qwen/Qwen3-235B-A22B-Instruct-2507	89.0	86.7–91.2	100%	307
9	Qwen/Qwen3.8-27B (stock)	88.7	86.5–90.8	100%	2,517
20	ibm-granite/granite-3.3-8b-instruct (stock)	86.4	84.4–88.2	99%	266
71	ibm-granite/granite-4.2-3b (stock)	64.5	59.7–69.0	80%	4,411

The full table is data/scoreboard.csv; per-field scores for every model are in data/per_field.csv.

What the arena showed: small fine-tuned models

The arena's leader before fine-tuning, Qwen/Qwen3.5-27B, became a trainer: it extracted the same fields from a large set of JOA listings, and a refinement step kept only answers grounded in the posting text (a title or a salary that actually appears in it). None of the 250 arena listings was in the training data, and the Claude-built reference labels were used only to evaluate, never to train. LoRA adapters on three open base models became JOA Small, JOA Advance and JOA Deep (status pages, no weights yet: https://huggingface.co/JobOpportunitiesAPI/joa-small, https://huggingface.co/JobOpportunitiesAPI/joa-advance, https://huggingface.co/JobOpportunitiesAPI/joa-deep).

• The smallest model ranks first. Fine-tuning lifted granite-4.2-3b from rank 71 to rank 1. JOA Small and JOA Advance are statistically tied; JOA Deep is just behind. Once the task was learned, size stopped mattering on this test.
• The students beat their teacher. Partly because the trainer sometimes reasoned until it ran out of tokens, and partly because the training answers were tidied into one consistent form. On the listings the trainer did answer, JOA Advance was still 1.9 points ahead (95% range +0.5 to +3.5).
• They answer in about 110–120 tokens. The trainer needs about 2,900, almost all of it reasoning.
• A stricter second test used 250 different, recent listings, each verified twice by Claude Opus against the live page and the application flow. On exact match, all three fine-tunes scored about 85%, the trainer about 86% and the stock base models 71–82% (2 October 2026). Here the teacher is still slightly ahead.

Known limits

• The models read what JOA stores, not the live page. The stored text is usually the description body; the header, where the title, location and posted date often sit, can be missing.
• They cannot tell that a job has closed. From stored text alone, all three called a few dead listings open. The fix is better input, not a bigger model.
• JOA Deep is weak on posted date (72, against 89 for JOA Small).
• JOA Small cannot drive a tool-using agent. It works when given the input, not when asked to fetch it.
• The references and the judges are models. Two judges, checked references and confidence ranges reduce that risk; they do not remove it.

Files

File	Rows	Content
data/scoreboard.csv	84	rank, model, kind (stock, gguf-quantised, fine-tuned), rule A score with 95% range, rule B and C scores, share of listings answered, median output tokens, size in GB for stock checkpoints, statistically tied with #1
data/per_field.csv	84	rule A score per field, rounded to whole points
data/gold_labels.csv	221	reference labels: arena_id, listing URL, host, the 11 fields, per-field confidence, per-field "findable in model input" flag, difficulty band

What is not included, and why

• The posting text. Job postings are written by employers and remain theirs. We publish the listing URL instead. Many of these postings have since closed, so the exact input cannot be rebuilt from the URLs alone; the scores above are the record of the run.
• Listings we do not share. The reference labels cover 221 of the 250 listings: those from employer-direct sources that JOA is allowed to redistribute. Left out: listings from a public employment service feed we use only for discovery, from job boards and aggregators, and from small employers named after a person. The scores were computed on all 250.
• Personal data. No names, email addresses or phone numbers of contacts. Free-text evidence quotes used while building the labels are left out.

Licence

The scores, per-field tables and this card are licensed CC BY 4.0. The reference labels in data/gold_labels.csv are provided for evaluation under the JOA Arena Terms in LICENSE.md: use them freely to evaluate, benchmark and publish results, including commercially, but do not use them to train or fine-tune models. The labels were produced with the help of Claude (Anthropic); this notice keeps the set an evaluation set. To help keep it out of training corpora, the files carry this canary string: JOA-ARENA-CANARY ac45c5b8-da08-4fb3-8e43-0c178dc94806.

Model names in the tables are those of their publishers; their mention does not imply endorsement.

Citation

Job Opportunities API (2026). JOA Job-Posting Extraction Arena, scoreboard of 2 October 2026. https://huggingface.co/datasets/JobOpportunitiesAPI/joa-extraction-arena

Acknowledgement

This work used the EuroHPC supercomputer Discoverer+ in Bulgaria, made available by the EuroHPC Joint Undertaking through an AI Factories Playground access allocation (project EHPC-AIF-2026PG01-1124).

Removal

Employers or individuals who want a listing removed: hello@jobopportunitiesapi.org. More about Job Opportunities API (JOA): https://jobopportunitiesapi.org

Contact

• General: hello@jobopportunitiesapi.org
• Legal matters: luca@tzekos.eu
• Phone: +30 2311 113 603
• Job Opportunities API (JOA) is a sole proprietorship of Loukas Tzekos, Didaskalisis Papathanasiou Vas. 79, 54629 Thessaloniki, Greece. VAT EL117613696.