From training to delivery

Define the task once

TaskRecipe connects the native model, data, objective, prediction and evaluation. The application owns their meaning, including batching and weighting. A prediction factory can construct TaskPredictor without rebuilding the training loop or reading training labels.

Prepare the declared inputs

Declare source records and files, then apply the application's deterministic preparation transform. A snapshot preserves source identity; a preparation cache reuses matching committed derived work. Decoding, collation and training-time augmentation remain application decisions.

Train, then continue the original work

The local training profile writes complete checkpoints with the supported optimizer, scheduler, RNG/scaler and data progress. A failed unfinished update does not replace the last complete checkpoint. Recipe recovery finishes the original target; requesting steps on an ordinary continuation means additional optimizer updates.

• Validation-selected inference assets and resumable training checkpoints have different roles.
• Data-order guarantees depend on the declared loader state.
• Qualified recipe recovery covers local Torch and single-process Accelerate, not arbitrary sharded or distributed continuation.

Commit prediction results

PredictionRunConfig preserves the request population and publishes checked structured records and declared files. Resume verifies and reuses completed parts without model replay. Work that never reached the commit boundary can require another callback invocation.

Evaluate what was actually produced

EvaluationSuiteConfig runs application-defined views over committed outputs and original references. It preserves coverage and per-record execution evidence alongside measurements and acceptance criteria. Missing references remain unavailable; evaluation does not invent targets or regenerate predictions.

Package the task and export its results

Package the selected model, inference resources and reconstruction code. A compatible consumer can invoke a declared local runtime entrypoint. ResultExportConfig separately formats committed results with their file closure and provenance; export does not replay prediction or submit the archive elsewhere.

Follow the installed tutorial

Generate a new task project first, then run these commands in that directory. They execute real application stages, including training. TUTORIAL.md adds interruption, committed-result resume and offline-consumer exercises to the same project.

python workflow.py prepare
python workflow.py train
python workflow.py package
python workflow.py predict
python workflow.py evaluate
python workflow.py runtime
python workflow.py export

Read the generated project

The project separates source assets, derived cache, training attempts and delivery artifacts. workflow.py is application code calling ARTI; project.json describes the selected paths and stage evidence.

my-task/
├── project.json
├── code/task_app.py
├── configs/
├── workflow.py
├── TUTORIAL.md
├── data/
│   ├── source/
│   ├── snapshot/
│   └── cache/
├── runs/
└── artifacts/