Technical Appendix
This appendix is not finalized. Specifications and limits are provisional. The appendix will be finalized no later than one month before the contest.
Updated: 21 September 2026.
This appendix applies to both the Individual Contest and the Team Contest. Each team receives the same resources and limits as one individual contestant.
1. Platform and Development Environment
| Component | Specification |
|---|---|
| Contest platform | Provides task statements, datasets, solution submission, final submission selection, and evaluation scores. |
| Operating system | Ubuntu 26.04 LTS |
| Main development environment | JupyterLab 4.6, accessed through the internal contest network, with GPU access on the training and evaluation machines. |
| Offline editor | Visual Studio Code (VS Code) 1.138 on contestant laptops, without direct GPU access. |
| Python | Python 3.13 |
2. Available Python Packages
The listed version families may change following installation and GPU compatibility testing. Final package versions will be published before the contest. Within each contest, all participants use the same software environment.
| Category | Package | Version family |
|---|---|---|
| Core AI/ML | torch | 2.14 |
| Core AI/ML | torchvision | 0.29 |
| Core AI/ML | torchaudio | 2.11 |
| Core AI/ML | transformers | 5.17 |
| Core AI/ML | accelerate | 1.15 |
| Core AI/ML | peft | 0.21 |
| Core AI/ML | trl | 1.13 |
| Core AI/ML | scikit-learn | 1.9 |
| Core AI/ML | xgboost | 3.4 |
| Core AI/ML | lightgbm | 4.7 |
| Core AI/ML | catboost | 1.2 |
| Core AI/ML | sentence-transformers | 6.0 |
| Core AI/ML | datasets | 5.0 |
| Core AI/ML | evaluate | 0.4 |
| Core AI/ML | spacy | 3.8 |
| Core AI/ML | nltk | 3.10 |
| Core AI/ML | gensim | 4.4 |
| Data processing | numpy | 2.5 |
| Data processing | pandas | 3.0 |
| Data processing | scipy | 1.18 |
| Data processing | polars | 1.44 |
| Data processing | pyarrow | 25.0 |
| Data processing | h5py | 3.16 |
| Computer vision | opencv-python-headless | 5.0 |
| Computer vision | Pillow | 12.3 |
| Computer vision | scikit-image | 0.26 |
| Computer vision | albumentations | 2.0 |
| Visualization | matplotlib | 3.11 |
| Visualization | seaborn | 0.13 |
| Visualization | plotly | 7.1 |
| Utilities/training | tqdm | 4.70 |
| Utilities/training | joblib | 1.6 |
| Utilities/training | tensorboard | 2.21 |
| Utilities/training | pytorch-lightning | 2.6 |
| Utilities/training | pydantic | 2.13 |
| Utilities/training | pyyaml | 6.0 |
| Development | jupyterlab | 4.6 |
The Python standard library may also be used. Installing additional packages during the contest is prohibited. TensorFlow and Keras are unavailable.
3. Hardware Resources and Pretrained Models
3.1. Laptops
Each individual contestant or team receives one Ubuntu laptop without a GPU. In the Team Contest, all three contestants share the laptop.
3.2. Training and Evaluation Machines
Training and evaluation use Amazon EC2 g6.xlarge instances:
| Resource | Specification |
|---|---|
| GPU | 1 NVIDIA L4 |
| GPU memory | 24 GB nominal (approximately 22 GiB) |
| CPU | 4 vCPUs |
| System memory | 16 GiB |
| Local instance storage | 250 GB NVMe SSD |
Each individual contestant or team has a dedicated GPU allocation for development.
3.3. Pretrained Models
The approved pretrained-model list will be published before the contest. Approved checkpoints and their required supporting files will be pre-cached in the contest environment. Contestants may use only models explicitly provided by the organizers. Documentation for approved models will be available offline.
4. Offline Resources
Required datasets and documentation are provided within the contest environment.
The tentative plan is to provide offline documentation through DevDocs for all listed packages. The final documentation collection and any additional offline references will be confirmed with the finalized appendix.
5. Submission and Evaluation
| Limit | Rule |
|---|---|
| Evaluation runtime | Maximum 10 minutes per evaluation run, unless the task statement explicitly states otherwise |
| Submissions | Maximum 15 submissions per task per individual contestant or team |
| Upload size | 100 MB hard limit per submission; tasks may set lower limits |
Notes:
- Multiple submissions may be queued concurrently.
- Runtime limits use wall-clock time and include training, model loading, preprocessing, and inference within the evaluation run.
- Submitted files are fixed at submission time; later edits do not affect them.
- Task statements specify submission requirements and evaluation procedures.