About 10.24

10.24 is a data science puzzle contest. It opens at 10:24 on 24 October and runs for five days. You can solve the problems in any order, with any language or tool you like.

How it works

• Typed-answer problems. Each one has a story and a question. Most give you your own input file: everyone's data is a little different, so answers can't be copied. Work out the answer on your computer and type it into the box. Many have a second part that opens once you solve the first.
• Upload problems. You get training data and make predictions for a test set, like a Kaggle competition. Upload a file and it is scored straight away against hidden answers.
• The model-training problem. The last problem is different: you train a small language model to solve maths and physics questions, on one GPU, within 48 hours that start when you press Start. A starter kit does the whole loop for you, so you can take part without having trained a model before.
• The picture. The 1024 bits on the home page start as noise. Every point you earn turns some of them into signal. Finish everything to see the whole picture.

Points

Each part of a typed-answer problem is worth a fixed number of points, shown on the problem list. There is no bonus for speed, but when two people have the same total, whoever reached it first ranks higher.

Upload problems give points on a sliding scale: a simple baseline earns 0 and the target score earns full points. During the event the leaderboard uses your score on the public part of the test set. When results are revealed it switches to the private part, so don't tune your model to the public score alone.

Wrong answers

You'll be told whether an answer is right straight away. After a wrong answer you wait a minute before the next try, and the wait grows if you keep guessing. Typing the same wrong answer twice doesn't cost you anything.

AI and other tools

You can use any tools you like, including AI assistants. We don't check. You'll get more out of the contest if you do the thinking yourself, and the problems are written to be fun to work through.

Computers and GPUs

Every typed-answer problem can be solved on an ordinary laptop. Some upload problems go further with a GPU, and the model-training problem needs one: at most 16 GB of GPU memory (a T4 is exactly right), see its page. Rent one when you need it:

• Featurize
• GPUHome

For the model-training problem, pick a card with 16 GB of memory (or a bigger one and stay within 16 GB). Stop the machine when you aren't using it, because you pay while it runs.

We reimburse GPU rentals up to an allowance we will announce before the event starts. Keep your receipts or invoices and send them to us after the contest ends.

Fair play

• One account per person.
• Don't post answers, inputs or code publicly until the event ends. Talking about ideas with friends is fine.
• Don't try to attack or overload the site. Accounts that break these rules can be removed.

Your data

We store your name (or the name you choose), your login method, your answers and uploads. If you log in with GitHub or Google we also store your email address, which only the organizers can see; we use it to contact prize winners.

Organizers

10.24 (Data Science Track) is run by Allen Xu (@allenxu09) and Henry Zhang (@caezium).

Questions

Find an organizer at school or email us.