About the project
Fiszka is an open topic classifier for AI papers and news: give it a title and a text, and it returns from zero to three of 17 topics, such as Agents, Security, Robotics or AI in Business. It is a 252,194-weight network on top of EmbeddingGemma 2 embeddings, small enough to compute in this browser tab.
Run it with Ollama and numpy: pull embeddinggemma-2:270m, then python fiszka.py --title "..." --text "...". The output is one JSON line per text with the chosen topics and a probability for each of the 17. No GPU and no cloud are needed.
Everything needed to rebuild it is public: 9,968 training texts with both teachers' answers and the vectors, and train.py, which reproduces the released weights exactly. Labels came from two open-weight models, DeepSeek V4.1 Flash and NVIDIA Nemotron 3.5 Lightning; a topic counted only when both gave it.
Against topics assigned by hand on texts it never saw, Fiszka scores F1 0.510 and 0.554 on two test sets, the same level as the labels it learned from. A bigger version 0.4 tied with it, so 0.3 shipped. Fiszka has tagged every new text in aipoc-lab since 7 November 2026.
Fiszka is open (Apache 2.0 licence): the model with its card and results on Hugging Face, the code on GitHub.
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)