Skip to content

Adaption Dataset Challenge

How the challenge works

Three weeks to build an instruction-tuning dataset with Adaption, train a model on it with AutoScientist, and show that it works.

Two crystal blocks side by side, one cloudy and cracked with flecks of pixels, the other clear.

In the Adaption Dataset Challenge, teams build an instruction-tuning dataset for an AI assistant serving an underserved language or domain. You define a use case, prepare a seed dataset, create or improve it with Adaption’s Invent a Dataset and Adaptive Data, train a model with AutoScientist, then evaluate it and submit your work. It runs fully online from Nov 16 to Dec 6, 2026.

The theme

Instruction-tuning dataset for an assistant in an underserved language or domain

Identify a practical assistant use case, prepare a suitable seed dataset, use Adaption’s Invent a Dataset and Adaptive Data to create or improve it, and train a model with AutoScientist.

Directions to consider

  • A question-answering assistant for an underserved language.
  • An assistant that explains a specialised topic or service in clear, accessible language.
  • A domain-specific assistant designed for a clearly defined group of users.

These are suggestions, not set tasks. Explain who the assistant is for, the problem it addresses and the behaviour you want to improve.

The team workflow

  1. Create and verify your Adaption account

    Every member creates an Adaption account at adaptionlabs.ai/app with the same email used to register for the challenge. The team leader then uploads one PDF of everyone's account screenshots to the Account Verification stage on Unstop by Dec 6, 2026, 11:59 PM IST.

  2. Define the use case

    Pick a practical use case, the people the assistant is for and the behaviour you expect from it.

  3. Choose a seed dataset

    Select or prepare a seed dataset. Public data or your own data both work, as long as you have the rights to use and adapt it.

  4. Build the dataset on Adaption

    Prepare the data in a format Adaption supports, then create or improve your instruction-tuning dataset on Adaption's platform.

  5. Use Invent a Dataset and Adaptive Data

    Use Invent a Dataset, Adaptive Data or both, as fits your use case. Record the methods, settings and iterations you used.

  6. Train with AutoScientist

    Train a model with AutoScientist, following the supported workflow and model options covered in Adaption's documentation and onboarding.

  7. Evaluate the result

    Evaluate the dataset and the trained model using the evaluation approach agreed with Adaption. Report your results and back up any improvement you claim.

  8. Submit your work

    Submit once, on Unstop or on your team page in the Hackyard app, with one link to your dataset, model details, evaluation results and project documentation. Submissions close on Dec 6, 2026, 11:59 PM IST.

Dataset guidelines

  • Source. A public dataset, or one your team creates, that you have the right to use and adapt.
  • Licence and rights. You must have the right to use and adapt the data, and you name its source and licence.
  • Structure. Instruction and response pairs, or chat-style data, as suits your workflow.
  • Format and size. Use a format Adaption supports. Formats, size limits and run limits follow Adaption's documentation and onboarding.
  • Privacy and safety. No personal, confidential, private or otherwise unauthorised data, and no unsafe or illegal content.

Onboarding and resources

Adaption will provide participants with the platform access and resources needed to build datasets and train models during the challenge. Participant onboarding with Adaption is on , 9:00 AM EST / 7:30 PM IST, and covers account setup, dataset preparation, Invent a Dataset, Adaptive Data and AutoScientist.

Read Adaption’s documentation and the welcome deck and quickstart. Hackyard runs participant support for the rest of the challenge and posts updates in the challenge WhatsApp group.

Teams

Prizes go to the top three teams, and every member of those teams gets the prize credits. Each member registers with their own email, because Hackyard shares it with Adaption for challenge-related account access.

Teams have 1 to 4 members, so you can also enter solo.

Fully online, for anyone 18 or over

There is no venue. You work from wherever you are, from Nov 16, 2026 to Dec 6, 2026, and judging runs Dec 7 to 13, 2026. Participants must be at least 18 years old to use Adaption. The challenge is fully online, so you can take part from anywhere, subject to applicable laws and Adaption's platform terms. See the full schedule or the prizes.

Registration is open.

The challenge runs online from Nov 16 to Dec 6, 2026, with onboarding on Nov 18 and submissions closing Dec 6 at 11:59 PM IST.

Questions? [email protected]