Add benchmark website (GitHub Pages)

[sevmag]

Hi @RasmusOrsoe!

This PR adds a static website for NuBench under docs/, ready to serve via GitHub Pages — a leaderboard-style page presenting the paper's results, in the spirit of other ML benchmark sites.

Live preview (deployed from my fork): https://sevmag.github.io/NuBench/

What's included

• Interactive leaderboard — one tab per task (energy, direction, track/cascade, vertex, inelasticity), each with a cross-dataset bar chart of a headline metric, the full per-dataset tables from the paper (best-model highlighting, bootstrap uncertainties, CC/NC splits), and the corresponding paper figures with captions underneath.
• Dataset cards — the seven datasets with procedurally drawn geometry art, key stats (events, strings/OMs, energy range, CC/NC ratio), and all four ERDA download links each.
• Models, getting started, citation — model overview with GraphNeT source links, download/extract/load instructions from the README, and copyable BibTeX.
• Dark/light theme, responsive, plain HTML/CSS/JS — no build step, no external dependencies.

Data provenance

The results in docs/data.js were parsed programmatically from the paper's LaTeX table sources (arXiv:2511.13111, tables 2–6) rather than transcribed by hand, including the bold best-model markup (with statistical ties) and the gaps (GRIT not trained on Flower S, DeepIce direction-only, no Hexagon Ice LE inelasticity). The figures under docs/assets/figures/ are rendered from the paper's own PDF figure sources.

To deploy

After merging: Settings → Pages → Deploy from branch → main, folder /docs. The site will appear at https://graphnet-team.github.io/NuBench/ (all in-site links in this PR already point at graphnet-team).

Happy to adjust anything — content, wording, colors, or where results/figures should live.

🤖 Generated with Claude Code

[sevmag]

Hey @RasmusOrsoe! I let Claude run loose and thought this would actually be a neat addition :) Let me know what you think!

[RasmusOrsoe]

Hey @sevmag!

Thanks for putting a bit of time into this.

We were actually hoping to provide a public leaderboard when we started the project. At the time, paperswithcode was the obvious choice, but when we were ready to submit (much later) paperswithcode was in an odd transition state after being acquired by Meta. We never got around to revisiting this - perhaps that is now.

Regarding your PR:

I like the visuals - in particular the landing page, dataset and model cards.

The leaderboard itself looks nice. In its current form, I see a few problems:

• The mechanics of submitting new models/results is missing; i.e. this is more like an elegant way of presenting the results of the paper than a public leaderboard.
• The bar plot shown on top of the tables is somewhat arbitrarily chosen and should either be removed or adjusted.

So - there is a choice to make:

a) If we want to provide a public leaderboard for the community to use, I'd strongly suggest finding existing portals that provide the mechanics for this. It improves discoverability and also puts the burden of maintenance onto someone else.

Or

b) if you're proposing to present only the results of the paper in a neat, concise way, I would suggest implement the following changes:

• Adjust the wording of the page so that it becomes clear that this is just a convienient way of exploring the paper.
• Remove the bar plot
• Rethink the order of the sections. Most people clicking the link might just want to get started with the datasets. E.g. (landing -> getting started -> datasets -> models -> results .. )
• Verify manually that numbers presented by claude is correct (i did not do this).

Hope you're doing well!

[sevmag]

I think the visualization alone is already quite nice, even if that means that one would need to update it manually to add new models to the leaderboard. Christian Locatelli is still working on the plotting scripts, so once these are done, we could add the functionality so that just uploading the predictions automatically updates the website. I'll ask Claude to implement your suggestions for the wording and reading flow!

Have you checked HuggingFace for hosting a leaderboard? I'm not too sure if it costs money for our purposes, but that could make it easier (also for the data download, I believe)