Property	Value
Architecture	Custom Causal GPT (Transformer)
Parameters	~159K
Precision	F32 (Full precision, no quantization)
Format	GGUF
Tokenizer	Character-level
Training Corpus	English Stories (Instructions)
Training Steps	~1,200 steps
Training Device	Intel HD Graphics (DirectML / CPU)

Training

This model was trained entirely from scratch using a custom PyTorch GPT implementation on a personal CPU/DirectML device — no pretrained weights were used. The training corpus consists of instructional English stories and creative text formatted as Instruction: ... \n Response: ... pairs.

Prompt Format

Instruction: your question or prompt here
Response:

Usage

With llama.cpp

./main -m model_text_instruction_char-F32.gguf -p "Instruction: tell me a story\nResponse: " -n 200

With Ollama

ollama run hf.co/DemonKing1234/casual-text-instruction-f32

With Python (llama-cpp-python)

python

from llama_cpp import Llama

llm = Llama(model_path="model_text_instruction_char-F32.gguf")
output = llm("Instruction: hi\nResponse: ", max_tokens=100)
print(output["choices"][0]["text"])

Limitations

This is a tiny experimental model (~7M parameters) trained for a short duration — it is not production-ready.
As a character-level model, it may produce occasional spelling errors.
Best suited for short text completions and creative writing experiments.
Does not have world knowledge or factual grounding.

License

MIT — Free to use, modify, and distribute.