Data API

Maturity: verified

Source module: rffm.data

Symbol	Contract
CanonicalBundlePaths	resolve canonical artifact paths under one root
CanonicalRFBundle	open one manifest and Arrow metadata table for split-owned signal access
CanonicalRFDataset	map-style PyTorch dataset with indexed ArrayRecord batch reads
CanonicalDatasetSource	selected registry name and canonical location
load_canonical_dataset_source	read only datasets.<name>.canonical_gcs_prefix
rf_collate	right-pad semantic samples without tokenizing them

CanonicalRFBundle.dataset() creates train, validation, or test datasets that share manifest and metadata identity while opening separate signal shards. CanonicalRFDataset exposes name, snapshot_id, split_info(), and fingerprint_digest(). It intentionally has no remote-download, normalization, patching, or training methods.

See Also

• Canonical dataset contract

• Semantic RF records