Leaderboard

Live: huggingface.co/spaces/CERN/colliderml-leaderboardSource: spaces/leaderboard/

Public scoreboard for the six benchmark tasks shipped in colliderml.tasks. Browse rankings, submit your own predictions, or reproduce someone else's to earn credits.

Layout

One tab per task. Inside each tab you get:

A table of the top 100 submissions (sortable, read-only). Columns include submitter HF username, all task metrics, credits earned for that submission, and whether it's a shipped baseline.
A Submit predictions accordion — upload a parquet file, the backend re-scores it against held-out truth, and any metric that beats the current best earns credits.
A Reproduce a submission accordion — enter a submission ID and upload your own predictions. If every metric lands within 2% of the original you earn 20 credits. This is the self-policing mechanism that keeps the board honest.

Task discovery

The task list is loaded from the installed colliderml.tasks package at startup:

python

import colliderml.tasks as _tasks
TASKS = _tasks.list_tasks()

Adding a new task in the library is enough to surface it in the UI — no Space redeploy beyond a pin-bump on the colliderml package in requirements.txt.

The Space falls back to a hardcoded task list if the colliderml import fails (e.g. during a first-boot partial install).

Authentication

Submission and reproduction both require a HuggingFace OAuth sign-in. The received token is forwarded as a bearer credential to the /v1/benchmark/<task>/submit and /v1/benchmark/<task>/reproduce/<id> backend endpoints. Browsing the rankings does not require a sign-in.

Credit economics

Action	Reward
Submit a new best on any metric	30–50 credits (task-dependent)
Add a new baseline that becomes the bar to beat	100 credits
Reproduce a submission within 2% tolerance on every metric	20 credits
Contribute a new task or channel	200 credits

Credit balances are shared across every entry point in the platform (the pip package, the Simulation Form Space, and this Space all draw from the same backend ledger).

Required Space secrets

Secret	Purpose
`COLLIDERML_BACKEND`	Backend URL; defaults to `https://api.colliderml.com`

Running it locally

bash

cd spaces/leaderboard
pip install -r requirements.txt
export COLLIDERML_BACKEND=http://localhost:8000
python app.py

Leaderboard ​

Layout ​

Task discovery ​

Authentication ​

Credit economics ​

Required Space secrets ​

Running it locally ​

See also ​