Clarification on AI coding assistants and cloud GPU compute with competition data

DaT Parkinson's Prediction Challenge
1

Hello, I’d like to clarify two points about the “no third-party tools that store/retain data” and “never share the data” rules:

  1. Is it acceptable to use an AI coding assistant (e.g., Claude Code, GitHub Copilot) to help write preprocessing/training code, as long as the raw NIfTI files and labels are never uploaded to or read by the assistant itself — it only writes code, and I run everything locally with the actual data myself?

  2. Is it acceptable to train a model using a cloud compute service like Google Colab, Kaggle Notebooks, or a rented GPU instance (e.g., RunPod/Lambda), where the competition data would be uploaded to a private notebook/instance that only I can access, used purely as compute infrastructure (not an AI/LLM service)?

Want to make sure my workflow stays compliant before continuing. Thanks!

1 Like
2

Hi there!

On both points, the relevant language is: "Never share, copy, or publish the data. For example, you may not use tools like Codex and ChatGPT that store or retain uploaded data, though you may download model weights and run models locally".

So the things to watch for are whether the challenge data is being retained by a third party, whether it’s accessible to anyone who hasn’t agreed to the competition rules, and whether every copy of it can be deleted at the end of the challenge. Some coding assistants read local files by default, so competitors should check what a given tool has access to. If you’re using cloud compute, any workspace, notebook, or storage holding the data should be private to you, with the provider’s terms not granting them rights to retain your uploaded content.

1 Like
3

Thanks for the clarification. I would like to confirm whether the following Kaggle-only workflow is compliant and prize-eligible.

Kaggle’s current Terms of Use state that licenses for all User Submissions are royalty-free, perpetual, irrevocable, and worldwide, including Personal User Submissions (Terms of Use | Kaggle). Its Privacy Policy also says deletion from active systems generally takes about two months and encrypted backups may retain data for up to six months (Privacy Policy | Kaggle).

Would this workflow still be allowed?

  1. Use a private Kaggle notebook with no collaborators and no public sharing.
  2. Do not upload the challenge files as a Kaggle Dataset or save them as notebook output.
  3. Download the DrivenData files directly into temporary session storage, use them only during training, and securely delete them before the session ends.
  4. Retain only generic code and trained model weights.
  5. Keep all scans, labels, UIDs, patient-level predictions, and images completely inaccessible to AI assistants.

Separately, may an AI coding assistant see only scalar validation metrics (for example, overall OOF log loss/AUROC) and non-data-specific software errors, while never receiving scans, labels, identifiers, patient-level predictions, plots, or images?

I want to ensure the workflow satisfies the requirement that every copy can be deleted and that the provider’s terms do not grant retention rights before transferring any competition data. Thank you.