OpenMLRL.github.io

Static site for the OpenMLRL collective built with Hugo. The homepage and Research page use local templates and styles; the hugo-theme-gallery submodule supplies fallback templates. GitHub Pages builds the site in CI and bundles the separately built CoMLRL documentation under /CoMLRL/.

Homepage and research

• layouts/_default/home.html defines the homepage, with introductory copy in content/_index.md.
• assets/css/custom.css controls typography, layout, and responsive breakpoints. Inter is hosted locally in assets/fonts/, with its license.
• data/publications.yaml contains the publications shown on the /research/ page.
• layouts/partials/collective-visual.html contains the SVG collaboration illustration. Its animation respects reduced-motion preferences.

Local development

1. Install the Hugo extended binary (already available via Homebrew).

2. Pull the theme if needed: git submodule update --init --recursive.

3. Start the hot-reload server:

   hugo server --port 1314

   The site becomes available at http://localhost:1314. Hugo keeps the build in memory, so no files inside public/ are written to disk.

To create a production build locally, run hugo --gc --minify. The output still lands inside public/, which is ignored by Git so it will not be committed by mistake.

CoMLRL docs

The Pages workflow publishes both this website and the CoMLRL documentation. It runs on website pushes, manual dispatches, and comlrl-docs-updated events sent by CoMLRL after documentation checks pass. Those cross-repository events use the existing OPENMLRL_TOKEN secret in CoMLRL, with Contents write access to this repository. Both repositories' workflow changes must be pushed to main to enable the complete flow.

Each run resolves one revision of CoMLRL's current main branch. Six parallel jobs reuse cached translations for Simplified Chinese, Japanese, Korean, Spanish, French, and German. Missing or changed pages are translated on the GitHub runner using Ollama and qwen2.5-coder:3b; no translation API key is needed. A cold build can take hours on CPU. Cached builds skip Ollama entirely when translations are current, and valid completed pages are cached even if a later translation fails.

The final job runs docs/scripts/build.py --multilingual --cache-only, validates all seven languages, and bundles the generated site under public/CoMLRL/. Only a successful build is deployed; translation failures leave the previous site available. Translations are kept in Actions caches and temporary artifacts, not committed to Git. Legacy files in static/CoMLRL/ are replaced in the final deployment artifact, so they cannot overwrite the multilingual build.

For a local multilingual preview, run python3 docs/scripts/preview.py --multilingual in the CoMLRL repository and preview this website with hugo server --port 1314.