Paper Translate

Paper Translate is a browser PDF viewer for scanned math papers. It shows the original page on the left and an English translation on the right, using either local Ollama models or qwen/qwen3-vl-32b-instruct via OpenRouter.

Current setup

• Frontend: React + Vite + pdf.js
• Backend: local Express server
• Models:
  • local Ollama: qwen3.5:0.8b, qwen3.5:2b, qwen3.5:4b
  • hosted OpenRouter: qwen/qwen3-vl-32b-instruct
• Math rendering: KaTeX

Requirements

• Node 20+
• npm
• Optional: an OpenRouter API key
• Optional: Ollama running locally

Installation

npm install

If you want hosted OpenRouter inference, create api.txt in the project root with this exact format:

"openrouter"="YOUR_OPENROUTER_KEY"

api.txt is ignored by git.

You can also use an environment variable instead:

export OPENROUTER_API_KEY="YOUR_OPENROUTER_KEY"

Run locally

Start the backend:

npm run dev:server

Start the frontend in another terminal:

npm run dev:web

Open:

http://localhost:5173

Usage

• Upload a PDF or page image
• Choose the source language: French or German
• Choose a model:
  • local Ollama for private/local usage
  • qwen/qwen3-vl-32b-instruct for the strongest hosted result
• Press t or click Translate Current Page
• Press Esc to leave split view
• Enable Show raw to inspect the exact model output and prompts

Notes

• Translation is page-based and manual
• The best hosted model currently configured is qwen/qwen3-vl-32b-instruct
• The right pane uses a readable flow layout based on model structure, not strict page reconstruction
• Formulas are rendered with KaTeX when possible
• A simple in-memory rate limiter protects public demos from light abuse

Quality checks

Run tests:

npm test

Build everything:

npm run build

Language evaluation helpers are also available:

npm run check:language -- --text "This is English."
npm run eval:pdf-language -- --pdf translation_samples/german_paper.pdf --source German --samples 5

GitHub

A CI workflow is included to run tests and builds on pushes and pull requests.

GitHub Pages is not configured here because live translation requires a backend and an API key. You can still host the frontend separately later if you point it at a hosted backend.

Hugging Face Spaces

This project can be deployed to a Hugging Face Docker Space.

Recommended approach

• Create a new Hugging Face Space
• Choose Docker as the SDK
• Add a Space secret:
  • OPENROUTER_API_KEY
• Push this repository to the Space

The included Dockerfile builds the frontend and backend together and serves the built app from the Express server on port 7860.

Why this works better than GitHub Pages

• the app needs a backend for translation requests
• the API key must stay server-side
• Docker Spaces support private secrets and server processes

Budget safety

If your OpenRouter key has a small monthly cap, you should still be careful with public demos.

Current lightweight protections:

• in-memory request rate limit per client/IP
• manual page-by-page translation only

For a stricter public demo later, consider limiting uploads or serving only sample PDFs.