Minionese LM

A tiny 1.35M-parameter conversational decoder-only Transformer trained entirely from scratch on an original Minionese-inspired dataset. It accepts short English or Minionese messages and replies in a playful, consistent project dialect.

Chat with the hosted model at notaminion.com →

This is an unofficial fan research project. It is not associated with, affiliated with, or endorsed by Minions or Illumination.

Model details

• Architecture: custom decoder-only causal Transformer
• Parameters: 1,354,560
• Layers: 4
• Hidden size: 160
• Attention heads: 5
• MLP size: 640
• Context length: 256 tokens
• Tokenizer: 512-token byte-level BPE trained from scratch
• Pretrained/base model: none
• Training: 5,000 batches and 9,981,697 sampled non-padding tokens
• Weights: safetensors
• Runtime: PyTorch CPU

The model was trained with MLX on Apple Silicon, but the uploaded weights run with the included PyTorch implementation on CPU-only Linux, macOS, and Windows systems.

Evaluation

On 400 held-out behavioral prompts, the final model achieved a 99.0% combined intent/style/clean pass rate and test perplexity of 1.116, compared with a 4.688 unigram baseline. Full machine-readable results are included in evaluation.json.

Usage

git clone https://huggingface.co/maddiedreese/minionese
cd minionese
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python inference.py "hello"

Expected greedy-decoding output:

Bello! Tulaliloo, amiko!

Python usage:

from inference import MinioneseRuntime

model = MinioneseRuntime(".")
print(model.reply("bello, como estas?"))

The canonical training and evaluation source is available at github.com/maddiedreese/minionese.

Intended use and limitations

This model is intended for education, experimentation, and playful short-form conversation. It is not an official translator or a general-purpose assistant.

The model supports conversational intents represented in its synthetic training dataset. It does not understand arbitrary English vocabulary, and compound or out-of-domain prompts can produce mismatched or malformed responses. Outputs may also repeat or terminate early because of the model's deliberately tiny size.

The training data does not use pretrained model weights, film subtitles, screenplay text, private messages, or proprietary Minions dialogue. It was programmatically generated from an original project dialect.

Hosted demo privacy

The separately hosted demo runs model inference on Railway. Typed chat messages are not stored. Optional voice interactions are processed by ElevenLabs, as disclosed in the demo interface. The web application and voice infrastructure are not part of this model repository.

License

The source code and model artifacts are released under the MIT License. This license does not grant rights to third-party names, characters, or trademarks.