atria_types defines the typed data contracts that flow between every module in Atria — from raw input files through storage, training, and inference. Every sample in the framework is a strongly-typed Pydantic model, not a plain dict or tuple.

The problem it solves

In typical ML pipelines, data representation is implicit:

# What is "sample"? A dict? What keys? What types?
for sample in dataloader:
    image = sample["image"]       # could be PIL, numpy, tensor?
    label = sample["label"]       # int? list? one-hot?
    annotation = sample["boxes"]  # list? array? what coordinates?

Schema drift — where training uses one format and a downstream consumer expects another — is a common source of silent bugs. atria_types eliminates this by making the schema explicit and enforced.

Key abstractions

Relation to other modules

• atria_datasets uses BaseDataInstance subclasses as the schema for storage rows.
• atria_models receives BaseDataInstance through the data pipeline and returns typed ModelOutput.
• atria_insights wraps instances for attribution computation.
• atria_types has no dependencies on other Atria packages — it is the lowest layer.