Minionese LM: A Small Language Model Trained from Scratch

I built Minionese LM as a playful, end-to-end language-model experiment. The tiny decoder-only transformer was trained from scratch on an original Minionese-inspired dataset, accepting short English or Minionese prompts and replying in a consistent project dialect.

The project uses no pretrained weights, film subtitles, or screenplay text. It is an unofficial fan research project, not an official translator or an Illumination product.

The included local artifact trained for 5,000 batches and 9,981,697 sampled non-padding tokens. On 400 held-out behavioral prompts, the reported intent/style/clean pass rate was 99.0%, with test perplexity of 1.116 versus a 4.688 unigram baseline. These results describe the project’s own evaluation.

Deliverables include the model, original dataset workflow, evaluation artifacts, open-source code, and a hosted demo.

Demo: https://notaminion.com Source: https://github.com/maddiedreese/minionese