Casual-TextTiny (800K Parameter GPT)
A custom, character-level Generative Pre-trained Transformer (GPT) language model built and trained entirely from scratch.
This model is designed for ultra-lightweight local deployment and learning experimentation. It is optimized to run on standard CPUs using a custom Numba-accelerated inference engine with full KV-caching.
Highlights
- Parameter Count: 801,408 (~800k)
- Model File Size: 3.2 MB in GGUF format (
Casual-TextTiny-F32.gguf)
- Trained For: Creative text/story generation and simple instruction-following.
Architecture
- Layers: 4 Block layers
- Attention Heads: 8 Heads
- Embedding Dimensions: 128
- Context Length: 16 characters
Prompt Format
To trigger instruction-following behaviors, wrap your prompt in this template:
