raphaelmerx/tulun

Transparent and Adaptable Low-resource Machine Translation, through LLM post-editing # Tulun Transparent and Adaptable Low-resource Machine Translation, through LLM post-editing 🎥 Demo video | 🖥️ Live demo (using Bislama) | 📄 Paper ## Local installation 1. Install Python dependencies: ```bash python -m venv .venv source .venv/bin/activate pip install poetry poetry install ``` 2. Setup credentials: create a .env file in the root directory and add the following: ```bash # for Google Translate (optional) GOOGLE_APPLICATION_CREDENTIALS=' .json' # for Gemini, can also use OpenAI / Anthropic / others, see docs.litellm.ai GEMINI_API_KEY=' ' ``` 3. Run the server: ```bash ./manage.py migrate && ./manage.py runserver ``` You can now configure your install (target lang, import glossary, etc.) at localhost. After that, you can access the translation interface at localhost. ## Deployment 1. Install Docker and Docker Compose 2. Setup credentials: create a prod.env file in the root directory and add credentials, similar to example above. 3. Run the server: ```bash docker-compose up -d ``` Access your server at localhost. Can be deployed behind a reverse proxy like Nginx. ## Evaluation For evaluations in the paper: see the eval folder README. For the in-app eval mode, upload your evaluation set at `/admin/translations/evalrow/`. Upon entering a sentence that is part of the eval set, the app will automatically switch to eval mode.