Jerzyk — Polish decision model and jerzyk-bench
Copyright 2026 Fundacja Trenify

Licensed under the Apache License, Version 2.0 (see LICENSE).

This project builds on the following third-party work:

- Laya multilingual (convaiinnovations/laya-multilingual) and the `laya` Python package — base model and
  fine-tuning tool for Jerzyk. Apache License 2.0. https://huggingface.co/convaiinnovations/laya
- mmBERT-base (jhu-clsp/mmBERT-base) — multilingual encoder inside Laya multilingual. MIT License.
- Bielik 11B v3.0 Instruct (speakleash/Bielik-11B-v3.0-Instruct), SpeakLeash & ACK Cyfronet AGH — wrote the
  generated court / administrative / accounting test documents and filtered the training data. Apache License 2.0.
- Qwen3 8B (Qwen/Qwen3-8B), Alibaba Cloud — wrote the generated training documents. Apache License 2.0.
- llama.cpp — local inference of Bielik and Qwen during data preparation. MIT License.

Data sources:

- SAOS — System Analizy Orzeczeń Sądowych (https://www.saos.org.pl): real, already anonymised Polish court
  rulings. Court rulings are official documents and are not subject to copyright (art. 4 of the Polish
  Copyright Act).
- samorzad.gov.pl — public notices of Polish municipal, county and state offices. Official documents
  (art. 4 of the Polish Copyright Act); personal data of natural persons was replaced with [OSOBA] / [ADRES]
  before publication.
- The customer e-mails in the jerzyk-bench test and dev splits were written with Claude (Anthropic) for
  evaluation only. They were not used to train Jerzyk; the dev split was used only to fit the three
  calibration temperatures (one number per question type).

Jerzyk is an independent project and is not affiliated with TypeSafe, Convai Innovations, SpeakLeash, Alibaba,
Anthropic or SAOS.