Poles Crowdsource a Million Photos to Train Polish AI Model Bielik

The creators of Poland's Bielik language model have launched a new civic project. The SpeakLeash Foundation, together with the Cyfronet Academic Computer Centre at AGH University, has launched the Obywatel Bielik (Citizen Bielik) initiative, which lets anyone submit their own photos to help build Poland's first AI vision model.

How the Collection Works

Participants submit photos of places, local cuisine, animals, historical figures and regional traditions characteristic of Poland, or add descriptions to photos already uploaded to the project's gallery. Organizers call each submission a "digital brick." The entire collection is meant to become an open, publicly available dataset that will benefit not only the Bielik team but other research projects as well.

Why Photos, Not Just Text

The SpeakLeash team's previous successes were in language. The Bielik and PLLuM models were trained on Polish texts, but global vision models, trained mainly on Western data, regularly misdescribe Polish realities, from dishes to architecture and landscapes. Obywatel Bielik aims to change that by supplying the model with data described by Poles themselves rather than by outside annotators.

Let's show Poland and the world what our culture looks like through our own eyes - organizers of the Obywatel Bielik project

Business and Institutional Support

The project has a sponsor in InPost, whose CEO Rafał Brzoska has for months publicly championed the development of Polish artificial intelligence. The National Digital Archive and Polska Press Grupa have also joined the initiative, providing access to additional photo archives and archival expertise for building the dataset.

The involvement of public institutions and private capital sets Obywatel Bielik apart from typical research projects funded solely by grants. The organizers hope this model of cooperation will let them gather data faster than in typical academic projects, where building large datasets usually takes years.

What This Means for Poland's AI Scene

Poland already has the Bielik and PLLuM language models, which, despite their domestic origin, have recently performed poorly in independent rankings, even losing out on tests of knowledge of Polish culture. A vision model built on domestic data is meant to answer that problem in a new area, images, where reliance on foreign datasets is even more pronounced than in text.

For companies and institutions that rely on image recognition, from media outlets to public administration, a locally trained vision model could mean fewer errors in describing Polish realities than tools built with the American or Western European market in mind.

What's Next

The organizers have not yet given an exact end date for the collection drive beyond the stated goal of one quarter, nor a release date for the vision model itself. The SpeakLeash team says the entire dataset, like Bielik's earlier text corpora, will be made publicly available, which they hope will enable similar civic datasets to emerge in other European countries.

Sources: INNPoland.pl (innpoland.pl), SpeakLeash (speakleash.org), Rzeczpospolita Cyfrowa (cyfrowa.rp.pl)