Tributo Official Algorithms

This monorepo contains official algorithm implementations for Tributo. Each domain package is an independently buildable, testable, versioned, and publishable Wheel. Third-party algorithms use the same public contracts and entry-point path.

Tributo Core owns Ray execution, resource allocation, retries, input leases, checkpoint transport, evidence, Tune, Bundle publication, and inference. These packages own model mathematics, framework-native hooks, executable contracts, and algorithm-specific delivery plugins.

Package	Algorithms and roles
classical	Random Forest, Extra Trees, Logistic/Linear Regression, MultinomialNB, PCA, KMeans, synchronous SGD, Isolation Forest
boosting	Ray Train XGBoost and LightGBM, ONNX/UBJ export, native flavor
tabular-torch	DNN, nnPU/uPU, PU prior and metric utilities
timeseries	Temporal convolution, LSTM, and GRU classification
catboost	Conditional distributed CatBoost ensemble
representation	Distributed tabular autoencoder
transformers-nlp	Pre-tokenized Transformer classification
graph-pyg	GraphSAGE and relational R-GCN
recsys-torch	Two-Tower and Jagged EmbeddingBag with All-to-All routing
multistage-torch	Distillation and pretrain-to-finetune
causal-core	Difference-in-means, DML, and IV
causal-discovery	Distributed PC stability discovery
causal-dr	Doubly robust/AIPW estimation
causal-xlearner	Five-stage X-Learner and batch CATE flavor
causal-dowhy	DoWhy estimation/refutation and GCM root-cause analysis

Development

uv sync --all-packages --group dev
uv run ruff format --check .
uv run ruff check .
uv run mypy packages
uv run pytest tests
uv build --all-packages

Package tags follow <directory>-v<semver>, such as classical-v1.2.0, so one package can be released without publishing unrelated Wheels.