maddiedreese/swedish-chef

Swedish Chef LM

A tiny decoder-only conversational language model trained entirely from scratch on an original mock-Swedish-chef-style corpus. It responds to short English prompts with playful, semi-comprehensible mock-Swedish speech and always ends with Børk børk børk.

This is an unofficial fan-research project and is not affiliated with or endorsed by Disney, The Muppets, or The Jim Henson Company. The Swedish Chef and The Muppets are third-party properties. The model and corpus should not be presented as real Swedish or as a Swedish translation system.

Model

• 1,354,560 trainable parameters
• Four causal decoder layers
• Model width 160; five attention heads; MLP width 640
• 256-token context
• 512-token byte-level BPE tokenizer trained from scratch
• Random initialization; no pretrained weights
• 5,000 training steps and 13.2 million sampled non-padding tokens
• Best checkpoint at step 4,250

Evaluation

On 192 deterministic held-out behavioral samples spanning 24 supported intents:

• Intent relevance: 100%
• Exact Børk børk børk closing: 100%
• Clean completion: 100%
• Test perplexity: 1.1688

At temperature 0.6, meaningful French, German, Spanish, or Italian sentences occasionally appear. The model is intentionally narrow and does not have general English knowledge.

Run locally

python3 -m venv .venv
source .venv/bin/activate
python -m pip install "git+https://huggingface.co/maddiedreese/swedish-chef"
swedish-chef

This opens an interactive chat. Type /quit to leave. For a one-shot response, use swedish-chef "hello".

Expected greedy output:

Ahoy hello, ve are readee-readee! Børk børk børk

On first use, the command downloads the small safetensors model from this repository and caches it through Hugging Face. Pass --model /path/to/files to use a manually downloaded copy. The self-contained runtime uses PyTorch on CPU and works on macOS, Linux, and Windows.

Training data and limitations

The 68,000-example corpus is overwhelmingly original and synthetic, with four short, source-tagged canonical dialogue observations. It covers 24 controlled conversational intents. High held-out scores measure interpolation within that narrow domain, not broad language understanding.