maddiedreese/minionese

This project is a small language model that talks back in Minionese, the playful gibberish language spoken by the Minions from the animated films. It is trained completely from scratch on an original dataset the author created, without using any pretrained weights, movie subtitles, screenplay text, or other copied material. The author is clear that Minionese has no fixed grammar and that this is an unofficial fan research project rather than a real translator or an official product. Under the hood, the model is a small decoder only Transformer with four layers, a narrow width, and a custom tokenizer built specifically for this task. It was trained on Apple Silicon hardware using a framework called MLX, but the finished model can also run on regular CPUs on Linux, macOS, or Windows, so people without a Mac can still try it. On a held out test set of 400 example prompts, the trained model produced an appropriate style and clean response 99 percent of the time. Using the project involves installing it as a Python package, then either training a new model from random starting weights yourself or using the model that comes bundled with the repository. Once set up, you can type an English or Minionese phrase and get a reply back in the same invented dialect, either through a single command or an interactive chat mode. There is also a hosted version of the model available online for people who just want to try it without installing anything. The model is intentionally limited. It only understands the kinds of short conversational phrases represented in its training data, and it can produce odd or broken replies to phrases outside that scope. This is presented honestly as an expected limitation of training such a small model from nothing, rather than a flaw to be embarrassed about. The project is aimed at people curious about training tiny language models from scratch or fans of the fictional language itself.