A micro application based on CroissantLLM and running entirely in the browser with WebLLM. Translates texts between French and English without sharing your content with the server.
- Fully private = No data ever leaves your computer
- Runs in the browser = No server needed and no install needed!
- Works offline
- Easy-to-use interface
This tool is built on top of WebLLM, a package that brings language model inference directly onto web browsers with hardware acceleration.
To run this, you need a modern browser with support for WebGPU. According to caniuse, WebGPU is supported on:
- Google Chrome
- Microsoft Edge
- All Chronium-based browsers
It's also available in Firefox Nightly, but it needs to be enabled manually through the dom.webgpu.enabled flag. Safari on MacOS also has experimental support for WebGPU which can be enabled through the WebGPU experimental feature.
In addition to WebGPU support, you need to have enough available RAM on your device to run the model (~3,5Gb).
You can check if your browser or your device support WebGPU by visiting webgpureport.
We use CroissantLLM, a 1.3B language model pretrained on a set of 3T English and French tokens, which runs swiftly on consumer-grade local hardware. This model was developed by CentraleSupélec, and it takes up 2.7 Gb of storage in the browser's cache.
To use CroissantLLM with WebLLM, we compiled the original model following the compilation process of MLC. We quantized the model into several formats (from q0f32 to q3f16) and you can find all the files of each quantized model on Croissant's Hugging Face repository. We choose to use CroissantLLM with the half precision model, because it had the best performance-to-memory usage ratio.
You can also use another model. To do that, you can compile your own model and weights with MLC LLM. Then you just need to update app-config with:
- The URL to model artifacts, such as weights and meta-data.
- The URL to web assembly library (i.e. wasm file) that contains the executables to accelerate the model computations.
- The name of the model.
You also need to change the custom prompt added before the user's text in the prompt file.
If you need further information, you can check the MLC LLM documentation on how to add new model weights and libraries to WebLLM.
Disclaimer: We utilize the chat version of CroissantLLM, specifically fine-tuned for chat interactions rather than translation purposes. Hence, you may encounter some limitations or inaccuracies in the translation. Additionally, as with any LLM, it's possible to encounter hallucinations or inaccuracies in generated text.
You can try it here. You can also run the project locally and contribute to improve the interface, speed up initial model loading time and fix bugs, by following these steps:
- NodeJS >= 20 - https://nodejs.org/
- NPM
If you're looking to make changes, run the development environment with live reload:
# Clone the repository
git clone https://github.com/numerique-gouv/croissant-translate.git
# Enter the project folder
cd ./croissant-translate
# Install dependencies
npm install
# Start the project for development
npm run devTo compile the React code yourself, run:
# Compile and minify the project for publishing, outputs to `dist/`
npm run build
# Build the project for publishing and preview it locally. Do not use this as a production server as it's not designed for it
npm run previewThis repository has a workflow that automatically deploys the site to GitHub Pages whenever the main branch is modified
This work is released under the MIT License (see LICENSE).