🏠 Smart-Home GPT-2

Turn natural-language smart-home commands into structured JSON tool calls — with a 124M model that runs entirely in your browser. No server, no cloud, nothing leaves your device.

🍴 Fork. Builds on barometech/smart-home-gpt2 by Pavel D. Popovich (barometech) — the original GPT-2 smart-home fine-tune, dataset and training pipeline. This repo continues it with in-browser inference, the v6→v9 cascade, long-context v14, and voice. See Acknowledgments.

🇷🇺 Описание на русском — в TUTORIAL.md.

"dim the living room lights to 20%"
   ↓  GPT-2 124M (smart-home fine-tune, in your browser)
{ "name": "dim_light", "arguments": { "room": "living room", "brightness_pct": 20 } }

▶️ Try it now

Nothing to install — the model runs in your browser via WebGPU and is cached after the first load.

• GitHub Pages: https://lifeart.github.io/smart-home-gpt2-tool/
• Hugging Face Space: https://huggingface.co/spaces/lifeart/smart-home-gpt2-tool

Built-in command presets, voice input (in-browser Whisper), and an fp16 / fp32 / q8 precision switch.

Why this project

• 🔒 Private by design. Inference runs 100% in the browser (WebGPU + transformers.js). No server, no cloud API, no data leaving the device. Weights stream once from the HF Hub and are cached.
• ⚡ Small and fast. A 124M-parameter model — ~330 MB as fp16, runs on consumer GPUs through WebGPU. Hosting is a static site (the live demos are free GitHub Pages / HF Space builds).
• 🎙️ Voice in any language. In-browser Whisper transcribes speech in 99 languages straight to English, then the model emits the tool call. No API.
• 🧩 Always-valid JSON. Constrained decoding masks the output to a JSON schema, so you get syntactically valid {"name", "arguments"} with correctly typed arguments — every time.
• 📚 123 functions out of the box, with a 4096-token context window (v14) so dozens of full tool schemas fit in one prompt.
• 🔌 Pluggable. The model emits a platform-neutral tool call; map it to Home Assistant, Zigbee2MQTT, ESPHome, Apple HomeKit, Tuya or plain MQTT — see INTEGRATION.md.

How it works

A single 124M decoder plateaus around 57% exact-match, so accuracy comes from composition and decoding tricks, not a bigger model:

• v6→v9 cascade — v6 predicts the function name, v9 (an arguments specialist) fills the arguments, a clean-gate picks the final call. Runs entirely in the browser.
• Constrained decoding keeps the JSON valid and typed.
• Enum value-snapping maps a loose value to the nearest registry enum ("gym" → "basement gym").
• v14-ctx4096 extends the context window to 4096 tokens for long tool lists, with no short-prompt accuracy tax.

Full design and rationale: ARCHITECTURE.md.

Quick start

Use the hosted demos above, or run the browser app locally:

git clone https://github.com/lifeart/smart-home-gpt2-tool
cd smart-home-gpt2-tool/web
npm install
npm run dev          # → http://localhost:5173/

To wire the output into a real smart-home stack, follow INTEGRATION.md.

Accuracy at a glance

It's a 124M model with a real, honest ceiling — the numbers are not oversold.

The in-browser app ships the 59.3% cascade; the 81.7% figure needs an external Llama endpoint and is research-only. Full tables, the 124M-ceiling analysis, dtype trade-offs, and how to reproduce the headline number: BENCHMARKS.md.

Make it yours

• Connect to a real home — INTEGRATION.md has recipes for Home Assistant, Zigbee2MQTT, ESPHome, HomeKit, Tuya and MQTT.
• Add your own functions — schemas live in web/tool_schemas.js and data/tool_registry.json. Names close to ones the model already knows work right away; brand-new functions need re-training (training/).
• Other languages — Whisper's task: 'translate' handles 99 languages automatically; nothing to change.
• Shrink / speed up — fp16 (default) and q8 ship in the browser; CPU/server quantization is covered in QUANTIZATION.md.

Documentation

Models

Streamed from the HF Hub on first load, then cached by the browser:

• lifeart/smart-home-gpt2-v14-ctx4096 — default, 4096-token window.
• lifeart/smart-home-gpt2-v9 — 1024-token window.

License

MIT — see LICENSE. Fork freely.

Citation

@misc{popovich_smart_home_gpt2_2026,
  title  = {Smart-Home GPT-2: in-browser tool-calling on a 124M model},
  author = {Popovich, Pavel D.},
  year   = {2026},
  url    = {https://github.com/lifeart/smart-home-gpt2-tool}
}

Acknowledgments

This project is a fork of barometech/smart-home-gpt2 by Pavel D. Popovich (barometech / Tekhnozhrets) — the original GPT-2 smart-home fine-tune, dataset and training pipeline, released under the MIT License (see LICENSE). The upstream repository is linked as a git remote in this fork.