SpTn-Model-only-1P
A minimal experimental generative model with exactly 1 trainable parameter.
SpTn-Model-only-1P explores how much generative behavior can be produced under an extreme one-parameter constraint.
Model
- Trainable parameters: 1
- Parameter type: float32
- Architecture: Character-level generative model
- Framework: NumPy
- Model file:
sp_tn_1p.npz - Status: Experimental
- Language understanding: No
How It Works
The model contains exactly one trainable parameter: w.
The parameter is trained using gradient descent.
During generation, the process is:
Input character → Fixed feature function → w → Character probabilities → Generated character
The generated text is produced directly from the trained parameter through the model's probability calculation.
The vocabulary and feature function are fixed and are not trainable parameters.
Download
Download the Model
You can download the model directly with:
wget "https://huggingface.co/KtiyaKK/SpTn-Model-only-1P/resolve/main/sp_tn_1p.npz?download=true"
Download chat.py
Download the chat program directly with:
wget "https://raw.githubusercontent.com/parhamtaheri453-crypto/SpTn-Model-only-1P/main/chat.py"
Then run:
python chat.py
Both files should be placed in the same directory.
Installation
Install NumPy:
pip install numpy
Quick Start
Download the model:
wget "https://huggingface.co/KtiyaKK/SpTn-Model-only-1P/resolve/main/sp_tn_1p.npz?download=true"
Download chat.py:
wget "https://raw.githubusercontent.com/parhamtaheri453-crypto/SpTn-Model-only-1P/main/chat.py"
Run:
python chat.py
Example:
You: hello
Model: helloenemts aonddsoc xroxoe
The generated text is experimental and is not intended to be meaningful natural language.
Load the Model
To verify and load the model:
python load_model.py
Expected output includes:
SpTn-Model-only-1P
Trainable parameters: 1
Model loaded successfully.
Training
The model can be trained with:
python save_model.py
This trains the single parameter and saves the resulting model to:
sp_tn_1p.npz
Architecture
SpTn-Model-only-1P contains exactly one trainable scalar: w.
The model uses a fixed character vocabulary and a fixed feature function.
Only w is updated during training.
The generation process uses the trained value of w to calculate character probabilities.
Important Limitations
SpTn-Model-only-1P is an experimental research project.
It is not a general-purpose language model.
The model does not understand natural language or the semantic meaning of the user's input.
Its output is character-level synthetic text.
The extremely small parameter count severely limits the model's capacity.
The project is intended to explore parameter efficiency and extremely low-parameter generative systems.
Roadmap
The project will investigate progressively larger parameter counts:
1P → 100P → 1K → 10K → 100K → 1M → 10M
The goal is to study how generative behavior changes as the number of trainable parameters increases.
Repository
GitHub:
https://github.com/parhamtaheri453-crypto/SpTn-Model-only-1P
Hugging Face:
https://huggingface.co/KtiyaKK/SpTn-Model-only-1P
Disclaimer
SpTn-Model-only-1P is an experimental project.
It should not be compared directly with modern large language models.
The project is intended for experimentation, education, and research into extremely low-parameter generative models.