Next
Benchmark challenges on Next

Find out which method really works

Set up a benchmark challenge in your browser. Teams prepare their method on GitHub, train it on your data and submit it. Every method is scored on the same holdout data, and a leaderboard shows what works best.

next.eyra.co
The Next platform: the workflow of a benchmark challenge
next.eyra.co
The leaderboard of the challenge
What is a benchmark challenge?

One question, many methods, one fair test

You pose a research question that can be scored, such as "who will have a child in the next three years?". Teams from anywhere build a method to answer it. Because every method is tested the same way, on data nobody has seen, you learn which approaches really work.

The question

A prediction task with a clear way to score the answers.

Training data

Teams download it to build and tune their method.

Leaderboard

Ranks the methods, with a tab for every metric.

1Create your challenge

Up and running in minutes

Create a project, add a benchmark challenge item, and you're ready to set it up.

Your colleagues can join as admins of the project, so you can organize the challenge together.

next.eyra.co
Adding a benchmark challenge item to a project
2Make it yours

Your challenge, your look

Add your logo, a header image, an introduction and the challenge rules that participants agree to.

Optionally add a privacy statement and a helpdesk page with your contact details. The default language is English, because challenges are usually international.

next.eyra.co
Settings: title, subtitle, logo and header image
3Build the participant journey

Three ready-made tasks

Prepare method links to your template repository. Download data gives teams your training data. Submit method collects their submissions until the deadline.

Submissions close automatically at the deadline, and participants see it in their own time zone. Running several phases? Add a submit task for each, each with its own leaderboard.

next.eyra.co
The Prepare method task, with the link to the template repository
4Make every method comparable

The template repository

Teams build their method in a copy of your template repository on GitHub. The template fixes how a method reads the data and writes its predictions, so every submission can run on the holdout data in exactly the same way.

  1. Copy the template

    Each team forks your template repository.

  2. Build the method

    They train and tune it on your training data, for example in Python or R.

  3. Submit a commit

    They submit a link to one commit on Next.

  4. Run on holdout data

    After the deadline, every method runs in the same Docker setup.

    Hidden from participants

How Eyra can help

Setting up a challenge for the first time? You don't have to do it alone.

Open source

Start from an example

The PreFer template repository is open source. Adapt it to your own data and metrics.

Advice

Design your challenge

Not sure how to set up a fair challenge? Eyra advises you on how to set up your challenge.

Read more about the service

5The participant experience

Clear steps for every team

Participants see your challenge, follow the steps and submit their method. Until the deadline, they can update their submission.

next.eyra.co
The three tasks as participants see them
Follow the stepsGet the template, download the data, submit.
next.eyra.co
Submitting a GitHub commit with a short method description
Submit a commitWith a short description for the leaderboard.
next.eyra.co
The leaderboard, with a tab per metric
See the resultsRanked per metric on the leaderboard.
6Score and publish

From submissions to leaderboard

After the deadline, export the submissions and run them on your holdout data. Upload the scores as a CSV file, and Next checks them against the submissions.

Then publish the leaderboard: open, with every team's name and code, or blinded, where teams only see their own name next to the scores.

next.eyra.co
Uploading the results of the submissions
Example

The PreFer challenge

In the Predicting Fertility data challenge (PreFer), teams predicted who would have a child within three years, using Dutch survey and register data. It ran on Next.

  • Training data from the LISS panel, in collaboration with Centerdata
  • Methods submitted from a shared template repository, in Python or R
  • Leaderboards at fixed moments during the challenge

Ready to run your benchmark challenge?

Create a free account and set up your challenge. Pricing depends on what you need, so get in touch to discuss your plans.