LLMcompoz

Compose a language model from scratch — and learn how it works.

LLMcompoz is a PySide6 desktop app (with a matching Qt-free CLI) that walks you through building a GPT-style language model from nothing: a tokenizer you can watch split text, a transformer you assemble block by block, a training loop that draws its own loss curve, and a generation panel where you scrub temperature and top-k and watch tokens appear one at a time.

Nothing is hidden behind a framework Trainer. Every core file is meant to be read top to bottom — the attention math, the training step, the sampling loop.

The six stages (tabs)

1. 📄 Veri — load text; see character/token counts and vocab size
2. 🔤 Tokenizer — char vs BPE; live view of a sentence splitting into ids
3. 🧠 Model — set layers/heads/embedding; live parameter count
4. 🏋️ Eğitim — live loss curve, step logs, checkpoints
5. ✨ Üretim — prompt in, tokens out, live; play with temperature/top-k
6. 📚 Öğren — concept notes explaining what happens under each stage

Layout

src/llmcompoz/
├── core/          # the actual ML — readable, framework-free
│   ├── tokenizer.py   # CharTokenizer + from-scratch BPE
│   ├── attention.py   # multi-head causal self-attention
│   ├── model.py       # GPT: embeddings -> blocks -> head
│   ├── trainer.py     # explicit training loop with callbacks
│   ├── sampler.py     # generation helper
│   └── config.py      # ModelConfig / TrainConfig dataclasses
├── gui/           # PySide6 UI (panels wire into core via callbacks)
├── learn/         # educational content
└── cli.py         # Qt-free llmcompoz command

Quick start

Windows (PowerShell)

cd C:\Github\LLMcompoz
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -e .
llmcompoz info
llmcompoz train .\data\tiny.txt --steps 500
llmcompoz-gui

Linux / CachyOS (Bash)

cd ~/Github/LLMcompoz
python -m venv .venv
source .venv/bin/activate
pip install -e .
llmcompoz info
llmcompoz train ./data/tiny.txt --steps 500
llmcompoz-gui

Status

v0.1.0 — scaffold: working core (tokenizer, model, trainer, sampler), CLI, and a six-tab GUI shell. Panels are placeholders to be filled in next.

License

MIT