The model, tagged “PLLuM‑12B‑nc‑250715,” was trained on what NASK called a “much better prepared and larger” corpus that includes texts from government websites, the Public Information Bulletin and the Science Library, all collected in line with Polish and EU law, researcher Agnieszka Karlińska said.
The updated system is available in a basic edition, an “instruction” edition tuned for multiple tasks, and a “raised” edition that adds safeguards.
The last combines human‑rated prompt‑and‑answer pairs to sharpen accuracy while blocking misuse.
NASK said prompt‑injection attacks now succeed in only two to three cases per 100 attempts, “significantly fewer than in other open models.”
The three versions can be downloaded for free from the Hugging Face platform after users submit a request form.
Piotr Pęzik, operational head of the HIVE AI project behind PLLuM, said the team avoids “bulk copying” of existing large models, instead generating synthetic data that is reviewed by humans.
That approach “helps PLLuM grasp Polish cultural context, answer more precisely and generate fewer unnecessary words,” he added.
Pęzik said HIVE will “soon” unveil a prototype citizen chatbot to collect prompts ahead of deploying PLLuM in the government’s mObywatel app, with further releases due in coming weeks.
PLLuM debuted in late February, when the Digital Affairs Ministry announced the HIVE AI consortium—led by NASK—to build Polish‑language models for public bodies, business, academia and, eventually, everyday users.
(jh)
Source: PAP