jerzyk-bench

A Polish benchmark for decision models (typed questions with calibrated answers) on incoming correspondence: court letters, administrative notices, accounting documents and customer e-mails, plus real court rulings and real public notices.

🇵🇱 Polski benchmark modeli decyzyjnych dla korespondencji: pisma sądowe, urzędowe, księgowe i e-maile klientów, plus prawdziwe orzeczenia i obwieszczenia urzędów. Pytania: rodzaj pisma, czy biegnie termin, pilność, dział.

Files

File	Documents	Source	Labels
sad_test, adm_test, ksieg_test	120 each	generated by Bielik 11B v3 from recipes	by construction, automatically checked
email_test	120	written by Claude (Anthropic) from recipes	by construction, rule-checked, rules from SPEC.md
sad_dev, adm_dev, ksieg_dev, email_dev	30 each	as above	as above
sad_real_test	89	real court rulings, SAOS 2024–2026 (already anonymised)	document type from SAOS metadata
sad_real_fragmenty_test	31	SAOS texts that start mid-text, with no header	reported separately (type often undecidable)
sad_real_dev	120	SAOS 2025–2026, disjoint from test	document type from metadata
adm_real_test	56	real public notices from 10 offices (samorzad.gov.pl), de-duplicated by template, personal data replaced with [OSOBA]/[ADRES]	labelled by Claude following SPEC.md, checked by two blind re-labellings and adjudicated; not human-reviewed

Each line: id, domain, split, state (document text), questions (typed questions in the TypeSafe Jev systemone format), expected (gold answers), source (where applicable), author, labels and note.

Questions

• rodzaj / dzial / kategoria_kosztu: choice from a fixed list.
• termin: yes/no. Does a deadline for the recipient's action run?
• pilnosc: score 0–3, from the deadline length counted from the letter date: 0 nothing to do, 1 no deadline or > 14 days, 2 8–14 days, 3 ≤ 7 days or immediately. E-mails have their own scale: blocked customer = 2; damage, threat of court or "today" = 3.
• E-mail: rezygnacja (terminates service or account; a single-purchase return does not count), zdenerwowany (upset; a threat of court or media counts).

Every rule and every decision taken while building the set is listed in docs/SPEC.md in the code repository.

Configurations on the Hub: sad, adm, ksieg, email (generated; test and validation = dev), sad_real (test, validation), sad_real_fragmenty, adm_real. Domain codes: sad court, adm administrative, ksieg accounting, email customer e-mail.

from datasets import load_dataset
ds = load_dataset("trenify/jerzyk-bench", "adm_real", split="test")

Personal data

• Real court rulings come from SAOS, where parties are already anonymised. Before publication we also replaced names of judges, prosecutors and court clerks with [SĘDZIA], [OSOBA] and [PROTOKOLANT] (scripts/anonymize_rulings.py).
• Real public notices are official documents published by the offices. Names and addresses of private persons were replaced with [OSOBA] and [ADRES]. Names of offices, their official contact details and parcel numbers were kept.
• Generated documents use invented names and companies. PESEL numbers, phone numbers and e-mail domains were masked.
• Check before publication. All 296 real documents were then read in full by two independent automated read-throughs (Claude Opus and Claude Sonnet) looking for remaining personal data; 11 leaks were found and masked (MANUAL_FIXES in scripts/anonymize_rulings.py). Names of authors of cited legal commentaries were kept, as published literature. This was not a manual human review.
• Anonymisation can still miss something. If you find personal data in this dataset, write to contact@trenify.org with the document id (please do not quote it in public discussions), and we will remove it.
• Data protection and copyright requests (for example from a person a document concerns): contact@trenify.org.

Metrics

Accuracy per question, mean over domains, ±1 accuracy and MAE for scores, Brier score for yes/no, ECE (15 bins), and latency. See scripts/evaluate.py in the repository. Do not train on the test split. Use dev for model selection and calibration.

Results

Generated test set (480 documents) — accuracy (%)

Question	Jerzyk 1.0 (322M)	basal-2.0-mini (1.5B)	Qwen3 8B	Bielik 11B v3
Administrative: urgency	77.5	61.7	42.5	54.2
Administrative: document_type	90.0	92.5	90.8	95.0
Administrative: deadline	82.5	79.2	84.2	83.3
Customer e-mail: department	93.3	88.3	89.2	88.3
Customer e-mail: urgency	79.2	74.2	70.8	80.8
Customer e-mail: cancellation	99.2	100.0	97.5	96.7
Customer e-mail: upset	91.7	90.8	77.5	85.0
Accounting: cost_category	97.5	95.0	85.0	75.8
Accounting: urgency	64.2	35.8	48.3	63.3
Accounting: document_type	94.2	91.7	95.8	95.8
Accounting: deadline	85.8	95.8	94.2	77.5
Court: urgency	80.0	42.5	38.3	54.2
Court: document_type	84.2	96.7	94.2	92.5
Court: deadline	93.3	85.0	83.3	82.5
Mean over domains	86.4	80.1	77.3	79.9

Real documents (89 court rulings + 56 administrative notices) — accuracy (%), 95% confidence interval in brackets

Question	Jerzyk 1.0 (322M)	basal-2.0-mini (1.5B)	Qwen3 8B	Bielik 11B v3
Admin. notices (real): urgency	58.9 (46–71)	37.5 (26–51)	48.2 (36–61)	32.1 (21–45)
Admin. notices (real): document_type	85.7 (74–93)	76.8 (64–86)	83.9 (72–91)	85.7 (74–93)
Admin. notices (real): deadline	73.2 (60–83)	76.8 (64–86)	62.5 (49–74)	82.1 (70–90)
Court rulings (real): document_type	98.9 (94–100)	93.3 (86–97)	100.0 (96–100)	100.0 (96–100)
Mean over domains	85.8	78.5	82.4	83.3

	Jerzyk 1.0 (322M)	basal-2.0-mini (1.5B)	Qwen3 8B	Bielik 11B v3
ECE, real documents (lower = better)	0.104	0.127	0.197	0.142
Median latency per document (all models on the same computer)	0.05 s	0.28 s	1.75 s	3.01 s

Models compared: Jerzyk 1.0 (calibrated); basal-2.0-mini (MLX 8-bit, native /v1/systemone); Qwen3 8B and Bielik 11B v3.0 Instruct (GGUF Q4_K_M in llama.cpp, zero-shot, answer chosen from option log-probabilities). Mean over domains averages the questions of each domain first, then the domains, so every domain counts equally.

Note. These results were measured on the benchmark text before anonymisation of the real documents (names of judges, officials and private persons replaced with placeholders such as [SĘDZIA] and [OSOBA]). On the published, anonymised text Jerzyk 1.0 scores 85.5% instead of 85.8% on real documents. The other models will be re-measured on the published text in a later update.

Caveats

• The generated sets use recipes, so their language is cleaner than real mail. Report real and generated results separately.
• The test documents for court, administrative and accounting were written by Bielik 11B, which may favour Bielik slightly.
• Urgency and deadline conventions are those of SPEC.md. A model that has not seen them is evaluated zero-shot.

About

Jerzyk and jerzyk-bench are published by Fundacja Trenify (Trenify Foundation), a Polish non-profit foundation. Issues and pull requests are welcome on GitHub. Contact: contact@trenify.org (also for security, personal-data and copyright reports).

Citation

@misc{jerzykbench2026,
  title        = {jerzyk-bench: a Polish benchmark for decision models on official correspondence},
  author       = {{Fundacja Trenify}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/trenify/jerzyk-bench}}
}

License

Apache 2.0. Court rulings (SAOS) and public notices are official documents (art. 4 of the Polish Copyright Act). Generated documents use invented names and data.