By joining and submitting to the UNICORN Challenge, you confirm that you have read, understood, and agree to comply with all rules and guidelines outlined herein.¶
General guidelines 📜¶
User Accounts and Team Formation
- Each participant is required to sign up with one verified account.
- Anonymous participation is not allowed, only verified Grand Challenge accounts are accepted to participate. Participants must accurately display their true names and affiliations (university, institute, or company), and country of residence.
- All participants must form a team, even if the team consists of a single member.
- Each participant can only be a member of a single team.
- Members from the same institution or group should form a single team.
- Teams are not allowed to split.
- Submission limits are defined at the level of a team.
Conflict of interest Individuals directly involved in organizing the challenge, companies sponsoring the challenge, and individuals providing data for evaluation are not eligible to receive prizes. Teams that include such individuals may still participate and be ranked, but their involvement must be clearly disclosed to all participants and visibly indicated on the public leaderboard.
Disqualification The UNICORN challenge organizers reserve the right to disqualify any participant or team at any time for unfair or dishonest practices.
Withdrawal policy Participants reserve the right to withdraw from the UNICORN challenge and forego any further participation. However, prior sub- missions can not be retracted, and published results will remain accessible.
Continued evaluation By submitting, participants agree to make their models available to the organizers via Grand-Challenge.org for future evaluation related to the challenge. To support this process, participants may be invited to update the preprocessing components of their Docker containers to ensure compatibility with new tasks or data. Evaluation results may be included in the UNICORN Challenge evaluation paper.
UNICORN Docker guidelines 🐋¶
- Only fully automated Docker containers are supported. Semi-automated or interactive methods are not allowed.
- Each team must comply with the compute limits and budget allocation. The algorithm Docker image must not exceed 20 GB in size. Details on the time limits for each task are available here.
- All submitted Docker containers will run in an offline environment without internet access. Any necessary resources (e.g., pre-trained weights, packages) must be included in the container before submission.
- Pre-trained models are allowed for foundation model training.
- Pre-trained adaptation methods are not allowed. Only the few-shots provided to the evaluation Docker may be used to train an adaptation method.
- Custom adaptation methods must be uploaded to the UNICORN Github Repository and will be public.
- Foundation model weights/ feature extractors are not made public, except for prize-winning teams.
Final test phase and prizes 🏆¶
The UNICORN challenge offers prizes for top-performing teams. To submit to the final test phase, it is required to submit a 2-3 page description of the training data and methodology, including details about:
- The datasets used during training, including quantity.
- Whether the data is publicly available, and if so, where they can be found.
- Model architecture used during training and number of parameters the model has.
- Training resources, including training time, cloud provider/ data center was used and where is it geographically located.
- For language tasks, if a system prompt is used, it must also be accessible to the organizers.
Open source To be eligible for prize money, all components must be open- sourced, except the foundation model weights and training data.
Prize money restrictions The competition is open to participants worldwide, except if you are a resident of North Korea, Crimea or any other place covered by Financial measures, Restrictions on goods EU sanctions. The aforementioned regulation is in place because of a requirement to conform to EU regulations concerning the provision of resources and transfer of finances (challenge award money, etc...).
Task coverage Participation in all tasks in the UNICORN challenge is not mandatory. However, for the final leaderboard, any missing task will receive a score of 0 or the minimum possible value for the relevant evaluation metric.