ml-std

The standard library for Brunost ML and RL tasks.

One pinned set of packages. The Judge's ML sandbox image installs exactly one release of ml-std; a contestant installs the same release and has the Judge's environment on their own machine:

pip install ml-std==2.0.0
python -m ml_std check      # "environment matches the Judge runtime"

On Linux x86_64 add --extra-index-url https://download.pytorch.org/whl/cpu to get the CPU build of PyTorch (the Judge runs CPU workers; the default PyPI wheel drags in CUDA libraries). macOS and aarch64 wheels are CPU-only.

Nothing else is importable inside the Judge (no network, no pip), so this package's dependency list is the complete answer to "what can I use?".

What is in it

The contest environment of IOAI 2025 (Python 3.12.7, requirements.txt), reduced to the ML-relevant packages and pinned at exactly their versions, so training for Brunost is training for the Olympiad:

Area	Packages
Deep learning	torch 2.7.1, torchvision 0.22.1, torchaudio 2.7.1, pytorch-lightning 2.5.2, timm 1.0.16, einops 0.8.1
Hugging Face	transformers 4.54.0, tokenizers 0.21.2, datasets 3.6.0, huggingface-hub 0.34.2, safetensors 0.5.3, accelerate 1.8.1, peft 0.16.0, sentence-transformers 4.1.0, sentencepiece 0.2.1, evaluate 0.4.4
Classic ML and data	numpy 2.2.6, pandas 2.2.3, scipy 1.14.1, scikit-learn 1.6.1, xgboost 3.0.2, lightgbm 4.6.0, catboost 1.2.8, statsmodels 0.14.4, numba 0.61.2, pyarrow 20.0.0, networkx 3.4.2
Images, audio, text	pillow 11.1.0, scikit-image 0.25.0, librosa 0.11.0, nltk 3.9.1, spacy 3.8.7
Plotting	matplotlib 3.10.0, seaborn 0.13.2
RL and testing	gymnasium 1.2.0, pytest 8.4.1, tqdm 4.67.1

Deviations from IOAI 2025, both forced by Python 3.13 wheels: scipy 1.14.1 (IOAI 1.13.1) and sentencepiece 0.2.1 (IOAI 0.2.0). Left out on purpose: GPU-only packages (bitsandbytes, diffusers), notebooks (jupyter, ipykernel) and API clients, which are not part of a judged solution.

ml-std	Judge runtime	Python
2.0.0	python-3.13-ml-v1	3.13

Releasing a new runtime

1. Change the pins in pyproject.toml, bump version there and __version__.
2. python -m ml_std catalog > src/ml_std/runtimes/<version>.json (keep the old files: earlier runtimes stay known).
3. Set RUNTIME to the new runtime name (python-3.13-ml-v2, ...).
4. Build and publish the wheel; build the Judge image with ML_STD_SPEC=ml-std==<version> and map the new runtime in BRUNOST_JUDGE_SANDBOX_IMAGES.

Tasks already published keep their runtime; new tasks default to the newest. python -m ml_std runtimes prints every runtime this release knows about, which is what the Judge's GET /v1/runtimes serves to the task editor.