FrugalSloth v0.3.3 [LIVE DEMO] BETA
Universal Edge AI Trainer & Inference Engine — 100% browser-native, zero backend, zero data leaves your device.
Train neural networks (MLP from scratch), fine-tune transformers via ONNX, and export production-ready models — all in your browser.
Download frugalsloth-v0.3.3-web.zip from the releases page, unzip it, and open index.html in any modern browser. That's it. No server required.
Or deploy the dist/ folder to:
- Cloudflare Pages — drag & drop
- GitHub Pages — push
dist/to a repo - Vercel/Netlify — connect repo
- Any static host — SFTP upload
# On Windows, after cloning:
npm install
npm run dist
# Output: release/FrugalSloth-v0.3.3-Portable.exenpm install
npm run dev
# Open http://localhost:5173| Feature | Description |
|---|---|
| MLP Training | Multi-layer perceptron from scratch with TF.js WebGL |
| ONNX Export | Export trained models to ONNX format |
| Transformer Fine-tuning | Import ONNX models (BERT, etc.), train classifier head |
| 100% Private | No data ever leaves your browser |
| Offline Capable | Works without internet after load |
| Static Deploy | Single ZIP drops into any host |
Browser (Your Device)
├── TF.js WebGL Backend — training + inference
├── ONNX Runtime Web — transformer inference + export
├── IndexedDB — model persistence
├── Web Workers — background inference
└── Vite + React + TypeScript — UI
No server. No API keys. No subscription.
frugalsloth/
├── electron/
│ └── main.cjs # Electron main process
├── public/
│ ├── icon.png # App icon
│ ├── coi-serviceworker.js # Cross-origin isolation for WASM threads
│ └── ort-*.wasm / ort-*.mjs # ONNX Runtime Web binaries
├── src/
│ ├── App.tsx # Main app (all tabs)
│ ├── terminal.css # Dual-theme design system
│ ├── types/frugalsloth.ts # TypeScript interfaces
│ ├── utils/
│ │ ├── csvParser.ts # CSV/JSON/JSONL parsing
│ │ ├── onnxExporter.ts # ONNX model export
│ │ ├── engineTemplate.ts # Standalone inference engine generator
│ │ └── indexedDB.ts # Model persistence
│ └── workers/
│ ├── inference.worker.ts # Background inference
│ └── transformer.worker.ts # Transformer pipeline
├── dist/ # Built static files (npm run build)
├── release/ # Electron builds (npm run dist)
└── package.json # Scripts + electron-builder config
| Script | Output |
|---|---|
npm run build |
Static files in dist/ |
npm run zip |
frugalsloth-v0.3.3-web.zip for hosting |
npm run electron |
Launch Electron dev mode |
npm run dist |
Windows .exe (portable) |
npm run electron:build:win |
Windows installer + portable |
npm run electron:build:mac |
macOS .dmg |
npm run electron:build:linux |
Linux .AppImage |
CSV: First row = headers. Target column = last column by default.
feature1,feature2,feature3,label
1.0,2.0,3.0,classA
4.0,5.0,6.0,classBJSON: Array of objects with consistent keys.
[
{"feature1": 1.0, "feature2": 2.0, "label": "classA"},
{"feature1": 4.0, "feature2": 5.0, "label": "classB"}
]JSONL: One JSON object per line.
{"feature1": 1.0, "feature2": 2.0, "label": "classA"}
{"feature1": 4.0, "feature2": 5.0, "label": "classB"}| Format | Use Case |
|---|---|
| JSON Weights | Reload in FrugalSloth, use in custom JS code |
| ONNX | Deploy to ONNX Runtime (Python, C++, mobile) |
| Engine JS | Self-contained <script> tag for any HTML page |
- v0.3.3 — Train tab empty state, Electron setup, static ZIP
- v0.3.2 — Cache busting, hard flush, colored buttons, tooltips
- v0.3.0 — Model versioning, auto-naming (name_v1, name_v2)
- v0.2.0 — Full UI overhaul, dual theme, docs tab
- v0.1.0 — Initial release, MLP training, ONNX export
Custom ONNX model upload requires local execution of the code. GitHub Pages deployment and Desktop App have WASM path limitations, pending to be fixed in the final release.
GNU AFFERO GENERAL PUBLIC LICENSE V3
Made with ❤️ for the Local AI Community by PacifAIst