TimeNet

Download and explore time-series datasets through one standardized format.

[!NOTE] This is a pre-release version and is subject to change. We are actively working on improvements around performance and integrations, and welcome community contributions.

Time-series data is fragmented. TimeNet standardizes it. Every dataset used to ship in its own shape, forcing teams to rewrite the same loading code again and again. TimeF replaces that with one shared format and one set of tools to find, download, and load any dataset the same way, whether it holds ECGs, accelerometer traces, or market prices.

TimeNet hands you the data and stops there. Training, inference, and modeling are up to you.

We're actively growing TimeNet: adding datasets, integrating time-series ML models, and building connectors to data processing libraries. Contributions in any of these areas are welcome.

Full documentation: https://docs.timenet.ai/

How it fits together

A connector turns a raw source into a manifest plus parquet and publishes it to a registry. The client reads the manifest from the registry and loads the data. Reading never runs connector code, so everything a consumer needs to interpret the parquet lives in the manifest.

• BaseConnector is the only contract a new data source must satisfy.
• TimeFDataset is the in-memory model a connector populates during convert().
• TimeFWriter serializes a populated TimeFDataset to disk.
• TimeFReader reads a TimeF version directory back into a TimeFDataset.

Components

The project is a uv workspace with two packages under packages/, plus the registry they read from and write to.

See the architecture guide for the full map, and the concepts page for the terminology.

Install

Requires Python 3.11 or newer (tested on 3.11 to 3.13).

uv add timenet            # core: TimeF format, reader/writer, registry client
uv add 'timenet[cli]'     # add the timenet console command
uv add 'timenet[torch]'   # add load_torch (PyTorch Dataset); works with any torch build

Once installed, the CLI is available as timenet. See Get started to load your first dataset.

License

TimeNet is released under the MIT License.

Dataset licenses

The MIT License covers TimeNet's own code, not the datasets it fetches. Each dataset keeps its upstream license. Check the license and source_url fields on a dataset's card to see what applies and where the data comes from. Some sources, such as PhysioNet, only grant credentialed access, so follow their terms when you download. See Dataset licensing for the full note.