# Datasets

A dataset is a versioned collection of [records](records.md). Its address is `org/name@version`, for
example `chengsenwang/tsqa@1.0.0`. The id is a HuggingFace-style `org/name` pair. The version is the
semantic version of the upstream source.

<figure markdown="span">
  ![Six small record signals under the header org/name@1.0.0, a versioned collection of records](../assets/figures/dataset-example.svg)
</figure>

## 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.

```python
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.

## Where a dataset lives

A [registry](../registry.md) serves a dataset. It hands the compiled manifests and parquet to the
[SDK](../client.md). A registry never runs connector code. It can be a local directory, an S3 prefix,
or a remote host. The output of [build](../build.md) is itself a valid registry. The same
`org/name` id resolves across all of them.