README.md

metadata

title: JuliaGPT
emoji: 🏛️
colorFrom: blue
colorTo: purple
sdk: docker
pinned: false
license: mit
short_description: JuliaGPT — experimental GPT in pure Julia
app_port: 7860

JuliaGPT

An experimental character-level GPT in pure Julia exploring minimal vocabularies inspired by ancient Greek scriptio continua.

Built from scratch with scalar autograd — no ML frameworks, just pure Julia.

Trained on Aristotle's Rhetoric and Euclid's Elements with a 28-character vocabulary (a-z + space + period).

API

OpenAI-compatible inference endpoint:

curl -X POST https://lisamegawatts-juliagpt.hf.space/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages":[{"role":"user","content":""}],"temperature":0.8,"max_tokens":128}'

Endpoints

Method	Path	Description
GET	/	Health check
GET	/v1/models	List models
POST	/v1/chat/completions	Generate text

Architecture

• 1 transformer layer, 16-dim embeddings, 4 attention heads
• Custom scalar autograd engine (Value type)
• Character-level tokenizer — 28 chars + BOS = 29 vocab
• KV cache for efficient inference
• block_size=256
• ~5,000 parameters

Links

• Model checkpoint
• Training data
• Source code