How-to guides

Task-shaped pages. Each one assumes you have published the dataset from the quickstart.

Working with a dataset

• Annotations: publish dense labels without rewriting signals.

• Selection and copies: publish a reference view or an independent copy.

• Direct access: read the objects without going through the library.

Training

• Training with Grain: adapt records to a Grain input pipeline.

• Training with PyTorch: adapt records to a map-style dataset.

• Read performance: which settings exist, and what was measured.

Storage backends

• Distributed writing: worker shards and root-last publication.

• Google Cloud Storage: generation-pinned cloud storage.

• Amazon S3: create-only publication and compare-and-swap.

• Mounted buckets: what a gcsfuse mount can and cannot do.