Browser-Based Yoda Translator

May the force be with your validation loss.

Live Demo: yoda.jpoehnelt.dev

An autonomous ML research project that trains a tiny 525K parameter encoder-decoder transformer to translate English to Yoda's Object-Subject-Verb (OSV) syntax. The resulting model is exported to ONNX and runs entirely client-side in the browser using a Svelte & Astro frontend.

🚀 Architecture & Stack

• Model: 525K Parameter custom Encoder-Decoder Transformer (PyTorch)
• Training: Automated agentic research loop (inspired by Karpathy's autoresearch) running on Apple Silicon (MPS)
• Inference Pipeline: ONNX Runtime Web (onnxruntime-web)
• Frontend: Astro + Svelte 5 (with Runes for reactive UI state)
• Deployment: Cloudflare Pages via GitHub Actions
• Data Source: sentence-transformers/all-nli (filtered for training pairs)

---

🧠 Part 1: Autonomous ML Research Loop

An AI agent reads program.md, modifies train.py, runs a training experiment, and commits or reverts based on whether the validation bits-per-byte (val_bpb) improved.

program.md  →  Agent reads instructions
train.py    →  Agent modifies architecture/hyperparams
prepare.py  →  Fixed data pipeline (never modified)
infer.py    →  TUI inference after training

Running the Agent Loop

Point your coding agent at this repo and say:

Have a look at program.md and let's kick off a new experiment.

The agent will autonomously:

1. Read program.md for instructions.
2. Modify train.py (architecture, hyperparams, optimizer).
3. Run training (e.g. 5-min budget).
4. Check val_bpb, commit if improved, revert if not.
5. Repeat.

---

🌐 Part 2: Browser Deployment

To make the model usable by anyone without a backend GPU, we export the PyTorch model to ONNX with embedded weights.

Exporting the Model

uv run python export_onnx.py

This generates encoder.onnx and decoder.onnx into web/public/models/. Because the models are extremely small (~11MB combined), they are tracked directly in Git to avoid long CI build times.

Running the Web Frontend Locally

The web frontend uses Astro and Svelte 5. It handles tokenization manually in JavaScript and streams inference via WebAssembly.

cd web
pnpm install
pnpm run dev

Deployment

The application is automatically built and deployed to Cloudflare Pages via GitHub Actions on every push to the main branch.

To deploy manually via the Wrangler CLI:

cd web
pnpm run build
pnpm exec wrangler pages deploy dist --project-name yoda-translator-1ml

---

📂 Project Structure

Directory/File	Purpose
prepare.py	Data loading, BPE tokenizer, evaluation
train.py	Model architecture & training loop
export_onnx.py	Traces and exports PyTorch to ONNX
infer.py	Terminal inference UI (rich)
web/	Astro/Svelte frontend application
web/public/models	Embedded .onnx files and tokenizer.json
.github/workflows	Cloudflare Pages deployment CI