A publishable entity in Atria is any artifact that carries enough information to be reconstructed by a consumer from the hub alone — without access to the original training code or local filesystem.

What makes an artifact publishable

Two things are required:

1. Content — the raw binary (weights bytes, storage shards, etc.)
2. Config / metadata — a serializable Hydra-instantiable dict that fully describes how to reconstruct the artifact

For model snapshots, this is the SnapshotArtifact.weights + SnapshotArtifact.metadata pair (see Snapshots & Hub).

For datasets, this is the dataset's storage shards + the serialized DatasetConfig (including preprocessing config, split info, and schema).

Model artifacts on the hub

A model artifact on the hub is identified by {owner}/{model-name} and optionally a branch name. It contains:

• model.safetensors — model weights in safetensors format
• metadata.yaml — the full ModelPipelineConfig + DatasetLabels in YAML format

Pulling a model artifact restores the exact ModelPipeline that was snapshotted:

from atria_models.core.model_pipelines._model_pipeline import ModelPipeline

pipeline = ModelPipeline.load_from_hub(
    name="my-org/resnet50-cifar10",
    branch="main",
)
# pipeline is identical to what was trained

Dataset artifacts on the hub

A dataset artifact contains: - The cached storage shards (Deltalake Parquet or msgpack) - The DatasetConfig JSON - Split info (train/val/test sizes, random seed)

Pulling a dataset artifact rehydrates a Dataset object identical to the one that was pushed:

atria datasets download my-org/cifar10-custom --output-dir ./data/

Versioning

Every push creates a new version. Previous versions remain accessible. The hub tracks: - Version number (monotonically increasing) - Commit hash (content-addressed) - Timestamp and pusher identity

This ensures that pulling a model or dataset by version always returns the same artifact, regardless of subsequent pushes.