Datasets¶

A dataset is a versioned collection of records. Its address is org/name@version, for example chengsenwang/tsqa@1.0.0. The id is a HuggingFace-style org/name pair. Dataset metadata separates the TimeNet release from the original source:

For example, a connector bug fix can produce TimeNet release 1.0.1 from the same source tag v2.1. The dataset_version changes to 1.0.1, while source_revision stays v2.1.

Versions are immutable¶

TimeNet commits a version atomically when its manifest.json lands. The version never changes after that. A pinned @version gives you exactly those bytes. With no suffix (or @latest), you get the newest committed version.

from timenet.client import TimeNet

client = TimeNet()
client.load("chengsenwang/tsqa")          # latest committed version
client.load("chengsenwang/tsqa@1.0.0")    # a pinned, immutable snapshot

Immutability makes full data lineage possible downstream. Every batch a model trains on traces back to the exact TimeF bytes of one version, not a moving target.

Parent dataset cards¶

A dataset card's parents maps local aliases to exact parent releases. A card declares each release under one alias only. The output manifest stores a copy of each parent's card in parent_cards. Each ParentDatasetCard holds that parent's metadata, whose dataset_id and dataset_version identify the exact release.

For example, the alias ptbxl can name physionet/ptb-xl@1.0.0. Its parent card records that release and its license, citation, access requirements, and other metadata. This copy lets readers inspect the build's provenance without loading the parent dataset.

Where a dataset lives¶

A registry serves committed TimeF versions to the SDK. It can be a local directory, an S3 prefix, or a remote service. A registry never runs connector code. A local build output is also a registry, so the SDK can read it directly.