Terms & Conditions

In order to participate in the competition, participants must register by accepting these Terms as described on the Competition website.

• Participants can work individually or in teams. Participants who work together on the same method are considered to be on the same team.
• A team is defined as one or more people who are jointly developing a solution to one or more tracks of the competition.
  • A participant cannot compete BOTH individually AND as part of a team.
  • A participant cannot be a member of multiple teams.
  • Prizes are awarded to the team members jointly.
  • It is the responsibility of participants to declare their team.
• Participants can work on any or all of the competition’s tracks (Weather, NMT, Motion Prediction). Participation in any track is not mutually exclusive of participation in other tracks.
• Participants should download the datasets from either the Data Provider’s website (WMT’20 sets for the NMT track) or from the Competition Website (Tabular Weather Data and SDC motion prediction data), and accept the applicable terms and conditions.
• Participants should build models only on the data provided for this competition by the Organizers. The use of any external data or pre-trained models for any of the three competition tracks is forbidden.
• Models must conform to the following computational requirements:
  • Models must be trainable on at most 8xV100 GPUs in under 16 days in real-time*.
  • Models must run on, at most, 1xV100 GPU and yield predictions within 200 ms per 1 input sample at inference.
• Participants should submit their predictions on the evaluation sets to be compared with hold-out references, which will be used to update the evaluation leaderboards.
• To be eligible for final scoring and prizes, participants are required to:
  • Submit a detailed 4-page technical report in the NeurIPS 2021 format on their solution.
  • Submit their code, which must be reproducible by the organizers.
  • Submit the model which corresponds to the submission.
  • Conform to the computational requirements.
  • Participants retain all rights to their code and submissions.
• Performance on evaluation data will be verified using the provided code, models, and technical report during the verification stage at the end of the competition. Only submissions that are verified to be consistent with submitted scores and which conform to computational requirements will be considered in the list of winning methods. Organizers will reach out to participants if difficulties arise.
• During the evaluation stage, we will limit the number of submissions on the evaluation data to 1 per day, for each track individually.
• In order to be eligible for prizes, the top ranked participants’ score must improve the baseline performance provided by the competition organizers.
• Yandex, OATML Group (Oxford), and CUED Speech Group (Cambridge) employees can participate in the competition, but are ineligible for prizes and will be excluded from the final leaderboard.
• Organizers have the right to update the rules in unexpected situations (e.g., a tie in rankings) or lead the competition in whatever way best meets its goals.

*For the Vehicle Motion Prediction track, if per-datapoint rendering is a significant bottleneck (e.g., if a pre-rendered representation cannot fit on a participant’s disk), the participant should provide utilities to benchmark their rendering time, which we will exclude from the 16 day timer. Rendering time can be defined as any compute time internal to iteration over a dataset (i.e., not including the forward pass, loss computation, and backward pass for a model trained via gradient backpropagation). For example, rendering time includes decompression, featurization, pre-processing transformations including rotations and down-/up-scaling.

For terms and conditions regarding prizes please see the following document.