What is Fiszka
Fiszka is the lab's own AI model: a small neural network that the lab built and trained from scratch on its own server. Only the model that turns text into numbers (EmbeddingGemma) is off the shelf. Every morning Fiszka reads the new AI texts the lab has collected and assigns each one from one to three of 17 topics, e.g. Agents, Security or AI in Business. Since November 2026 Fiszka assigns the topics on the blog; Gemini only tags alongside, for comparison.
It turns the title and body of a text into 768 numbers describing its meaning (a vector from the EmbeddingGemma model). These numbers pass through three layers of neurons (256, 128 and 64) and a shortcut straight to the topics. At the end comes a confidence for each topic, and a topic passes from 0.50. In total that is 252,194 parameters. Fiszka runs on the lab's server, without the cloud, and on this page your browser computes the same numbers.
It learned from 9,968 texts labelled by two large language models, DeepSeek and Nemotron. A topic went into training only when both agreed.
We checked it on texts the lab's author labelled by hand. Agreement with the human judgement (F1 score): Fiszka 0.51, Gemini 0.34 – in our own test, with our prompt for Gemini.
Fiszka is open (Apache 2.0 licence): the model with its card and results on Hugging Face, the code on GitHub.
Project page with the network running in your browser: ch3sh1r3xx11.github.io/fiszka.
Posts about Fiszka
- Mały model kontra Gemini: pierwsza próba otwartego klasyfikatora tematów AI (Fiszka 0.1, 4 Nov, in Polish)
- Mały model dorównał Gemini w tagowaniu tematów. Fiszka 0.3 (6 Nov, in Polish)
- Fiszka 0.3 is public. The model, model card, and a week in production alongside Gemini (14 Nov)