Application for compute through the Fields Model Initiative
This application will walk you through the steps to obtain GPUs to support your AIMO3 submissions. Only AIMO3-related efforts will be allowed, and you need to have registered on the AIMO3 Kaggle website (http://kaggle.com/competitions/ai-mathematical-olympiad-progress-prize-3 ) in order to be eligible for this compute. Please also read our technical submission instructions (https://github.com/hellofieldsmodel/submissions-instructions), more information on those below.
Sign in to Google to save your progress. Learn more
Email address of the main point of contact: Please provide a single email address of the main person who we will be in touch and who will act as team representative (if a team consists of multiple people). We will use this email address to contact you and by providing this email address you commit to be reachable and answer all emails within 16 hours of them hitting your inbox. *
Kaggle data 1: Please list the AIMO3 team name. *
Kaggle data 2: Please list all the members of your AIMO3 team.

Required format: every team member on a new line, starting with their name and all affiliations comma-separated after the name.

Example:
 *
Team Affiliations *
Please list all individuals who have significantly contributed to any submitted artifact that you will use, including but not limited to code, datasets, model training, data processing, and evaluation.
This list must include every person whose work is directly or indirectly incorporated into the submitted materials. E.g. if you use our GPUs to benchmark a base model that you finetuned, you must list who was involved in finetuning the model, and who was used in creating the benchmark.

The persons involved may go beyond the ones listed on your Kaggle team, in cases where, for example, you are part of a company and a subset of people compete in Kaggle, and other company members have created that in-house benchmark.

Required Format: every team member on a new line, starting with their name and all affiliations comma-separated after the name.

Example:
John Foo, University of Oxford, emailofjohn@somemail.com
Peter Bar, Technical University of Vienna, Some-Company, emailofpeter@somemail.com

Resource Requirements *
Please describe your Estimated Compute Requirements: Provide an estimate of the total GPU compute required for the project, expressed in GPU hours. You can assume NVIDIA H100/H200 GPUs as the standard reference. If possible, briefly justify your estimate (e.g., epochs, model size, training runs, etc.) You can request as much as 128 H100 GPUs, but the more you request, the more well formulated your submission needs to be. Our aim is to support everyone who submits, but since our process is not fully automatic, and we have to manually copy containers, we will prioritize best submissions first (more on that below).
Name all applicable licenses for any artifacts you submit *

To ensure compliance with AIMO3, all submitted materials must fully comply with applicable license agreements and legal requirements governing their use in scientific research and model development. This obligation applies to all artifacts involved in the project, including but not limited to:

  • Source code and scripts
  • Training frameworks
  • Pre-trained model weights, if applicable
  • Datasets (raw, processed, augmented, ...) in the form in which they are used
  • Evaluation benchmarks and other resources
You must not submit, reference, or use any artifact unless you hold a clear, permissible, and verifiable license that explicitly allows its intended use (e.g., research use, modification, redistribution, and/or derivative works, as applicable).

For each of the artifacts used by your submissions, you required to provide:

  • The exact license and version
  • Intended use of the artifact in the submission (e.g., SFT, Benchmark, ...)
  • Relevant restrictions or obligations imposed by the license (e.g., attribution, non-commercial clauses, ...)

All licenses must permit scientific reproducibility. At a minimum, this includes the right for researcher to:

  • Inspect and study your artifact
  • Reproduce any experimental results using the same materials
  • Verification of any claims made in publications, reports or social media
Artifacts that prohibit reproducibility, verification or prevent the reuse for research purposes are not permissible unless explicitly justified and approved.

The license information must be clearly documented for all artifacts to ensure long-term reuse and replication of the research outputs enabled by the FMI.

Response Format:
Each artifact must be listed on its own line using the following comma-separated structure:
<Artifact Name>, <License>, <Intended Use>, <Additional Notes (if applicable)>
  • Use one line per artifact. 
  • Fields must be comma-separated. 
  • Additional notes are optional but recommended where restrictions or obligations apply

Example:
Olmo 3 7B Instruct, Apache License 2.0, pre-trained model weights used as the base for SFT
NVIDIA OpenMathReasoning, CC BY 4.0, supervised fine-tuning (SFT) dataset, required attributions provided

Please note that since your container runs on an academic research cluster, we will not have capacity to debug the scripts you have containerized them or run them multiple times if something goes wrong the first time.

To ensure disruption is minimal, we therefore require you to rent public compute simply to test whether your Singularity container actually runs and does what it is supposed to do, before sending it to us. This does not need to be costly, as the main training run will be carried out on the cluster maintained by the LLMC/NII, and the role of this would mainly be to remove simply training bugs on multi-GPU system that your script may have.

There are public vendors out there, e.g. https://rescale.com/documentation/main/rescale-advanced-features/running-your-custom-code-on-rescale/using-apptainer-singularity/ seems to support Singularity (we are not affiliated with rescale.com and this solely for your information).

We have written very detailed and lengthy submission guidelines, available at https://github.com/hellofieldsmodel/submissions-instructions/, which also contain information on how you need to document your submission. 

Please provide below an Amazon S3 presigned URL. The meaning and what this URL should point to is described in the linked submission instructions.
 *
Submit
Clear form
Never submit passwords through Google Forms.
This content is neither created nor endorsed by Google. - Terms of Service - Privacy Policy

Does this form look suspicious? Report

Google Forms