Gadsby-4B

Gadsby-4B

Chats in a quirky chat-speak voice after a ban on word pairs common in Simple Wikipedia. A 4.3B-parameter Gemma 3 experiment for people curious about constrained decoding and how it changes a model's voice.

Made by errrinn on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Gadsby-4B is an experiment by errrinn, built on Google's gemma-3-4b-it. The author banned every token pair found in Simple Wikipedia during decoding, which pushed the model into a chat-speak voice with misspellings and a strong personality. A second step fine-tuned the model on its own constrained replies, keeping only the coherent ones.

Who it suits: People who are curious about how a limit on word pairs changes a model's voice, and hobbyists who enjoy an odd chat persona.

Industry | Who uses it | How they use it | Example
Research | Language-model hobbyists | Study how a decoding limit changes a model's voice | Comparing its replies with the base Gemma model
Creative work | Writers and game makers | Borrow the odd chat-speak voice for a quirky character | Drafting a goofy chatbot persona
Education | Machine learning teachers | Show constrained decoding and rejection-sampling fine-tuning | Walking a class through the two-step recipe

Anyone who needs clean, correct answers can skip it, since the card's own samples are noisy.

Where to use it:

• Windows, macOS or Linux with Python, for downloading the files
• Local tools that run Gemma 3 4B models, since this one is based on google/gemma-3-4b-it

Use case | Who | What they get back
Voice experiments | Researchers | Replies in shouty chat-speak with misspellings
Persona prototyping | Creators | A sample of a quirky assistant voice
Teaching | Machine learning teachers | A concrete case of constrained decoding

How to set it up:

• Python on Windows, macOS or Linux
• The huggingface_hub package
• Room for the weights, about 8.6 GB in BF16, or about 2.2 GB at 4-bit as an estimate

1. Install Python if you do not have it yet.
2. Install the huggingface_hub package with the first command below.
3. Download the model files with the second command.

pip install -U huggingface_hub
hf download errrinn/Gadsby-4B

How to use it: The card shows no run command, only chat samples. Because the model is based on google/gemma-3-4b-it, you would most likely load the downloaded files in a tool that runs Gemma 3 models and chat the usual way.

Example: A user writes "explain photosynthesis to me." The reply comes in shouting capitals with misspellings and chat-speak, and it says plants make food from sunlight, carbon dioxide and water, giving glucose and oxygen.

For newcomers and experts: Think of it as a model that was kept from using the usual word pairs of Simple Wikipedia and ended up sounding like an excited text chat. A good first step is to read the card's four chat samples and see the voice for yourself.

The recipe has two stages. Constrained decoding bans the token pairs, and rejection-sampling fine-tuning then trains on the model's own constrained outputs, keeping the coherent ones. The name nods to an older novel written without the letter e.

Ban Simple Wikipedia token pairs
Generate with the ban in place
Keep only the coherent outputs
Fine-tune on those outputs
Chat in the new voice

Good to know:

• License: CC BY-NC 4.0
• Cost to run: none named
• Data it reaches: the download fetches files from Hugging Face, and the source names nothing else
• Not for: commercial use, because the license is the non-commercial one
• The card's own sample answer to 17 times 23 is wrong, so do not rely on it for facts

On Hugging Face, 13 November 2026

• Parameters: 4.3B
• Weights: about 8.6 GB in BF16; about 2.2 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 90
• Likes: 15
• License: cc-by-nc-4.0
• Based on: google/gemma-3-4b-it
• Quantized versions: 1
• Tags: safetensors, gemma3, constrained-decoding, rejection-sampling, lipogram

Links

• Model: https://huggingface.co/errrinn/Gadsby-4B
• Files: https://huggingface.co/errrinn/Gadsby-4B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download errrinn/Gadsby-4B

Then follow the usage in its model card. Nothing in this kit runs by itself.