Skip to content

README.en.md

Latest commit

 

History

History
119 lines (88 loc) · 4 KB

README.en.md

File metadata and controls

119 lines (88 loc) · 4 KB
Raw

ChiefUI Assistant

English | Türkçe

A local, from-scratch AI assistant focused on UI/CSS generation. No pretrained LLM / cloud API is used (no LLaMA, Qwen, Ollama, GGUF, OpenAI, Claude). The model is trained from scratch in PyTorch.

Hardware target: NVIDIA CUDA GPU (~6 GB, e.g. GTX 1660 Ti / Turing). CPU fallback works for testing. MVP is single-PC; the architecture is ready to add a 2nd PC as a worker later (see docs / wiki).

Prompt → structured @@HTML / @@CSS / @@NOTES output → live iframe preview + code tabs + ZIP export.

📖 Wiki — Architecture · Training · Dataset Format · API

Architecture

Docs: docs/architecture.md · data: docs/dataset-format.md · training: docs/training.md

chiefui-assistant/
├─ backend/    FastAPI + PyTorch model (from-scratch GPT) + tokenizer + SQLite
├─ frontend/   React + Vite + Three.js (prompt, live preview, code tabs, ZIP)
├─ scripts/    train_tokenizer · prepare_dataset · train_model · generate
└─ docs/

The model is a small decoder-only Transformer built from scratch: RoPE positions, RMSNorm, SwiGLU MLP, (G)MHA + SDPA, KV-cache, trained with fp16 + gradient checkpointing to fit 6 GB.

Setup

1) Backend (Python)

cd chiefui-assistant

# Install torch for your CUDA (GTX 1660 Ti / Turing -> cu126 verified):
pip install torch --index-url https://download.pytorch.org/whl/cu126
pip install -r backend/requirements.txt

The same wheel runs on both GPU and CPU. No GPU? Use --device cpu.

2) Frontend (Node)

cd frontend
npm install

Run (MVP, end-to-end)

From the project root (chiefui-assistant/):

# 1) seed data (8 balanced categories) — skip if you have your own data
python backend/datasets/build_seed.py

# 2) train our own tokenizer
python scripts/train_tokenizer.py --input backend/datasets --vocab-size 16000

# 3) JSONL -> token bins
python scripts/prepare_dataset.py --input backend/datasets --out backend/datasets/bin

# 4) train (quick try: tiny + few steps)
python scripts/train_model.py --preset chiefui-tiny --data backend/datasets/bin --max-steps 1000
#   real run: --preset chiefui-30m  (takes longer)

# 5a) start the backend
cd backend
uvicorn app.main:app --port 8000
#   (model loads from checkpoints/ckpt.pt + checkpoints/tokenizer.json)

# 5b) frontend in a separate terminal
cd frontend
npm run dev
#   -> browser: http://localhost:5173

Test from the terminal (no backend needed):

python scripts/generate.py --instruction "Create a responsive hero section" \
  --input "Style: dark, orange #c45a26"

API

  • POST /api/generate-ui → { id, html, css, notes, raw, validation }
  • WS /ws/generate → {type:"token",text} … {type:"done", html, css, notes, validation, id}
  • GET /api/history?limit=20 · GET /api/health

Presets

preset ~params note
chiefui-tiny ~7–10M quick smoke test
chiefui-30m ~27M recommended for 1660 Ti
chiefui-60m ~50M larger, slower

Turing (1660 Ti) uses fp16 (auto-selected; bf16 is emulated → disabled).

Honest expectations

A from-scratch ~7–60M model works with small data but the output is basic at first — it picks the right component category and produces valid HTML/CSS, but design variety is limited to the seed templates. For high-quality, original generation you need thousands of good examples + longer training. The goal of this MVP is a working end-to-end pipeline and a clean, extensible architecture. Growing the dataset is the highest-impact step (extend build_seed.py or add your own *.jsonl).