Hebrew Codenames
A semantic-alignment experiment using Hebrew word embeddings, model comparisons, and an interactive map.
Open the original ↗
Comparing meanings through a word game
The project uses Codenames to compare how Hebrew encoders and language models interpret word associations. An encoder can suggest a clue from vector geometry; a language model can supply intended targets or rank guesses. Comparing the two exposes where their representations agree and where they diverge.
The repository distinguishes that agreement from clue quality. It includes a notebook-based probe, a local interactive map with DictaLM, and a static map that runs without an LLM.
Related talk
My Reversim Summit 2026 session, “How to Cheat at Hebrew Codenames (Without an LLM),” focuses on word vectors and geometry. The event page lists December 1, 2026 as the session date.
