gradio-app/trackio

A lightweight, local-first, and free experiment tracking library from Hugging Face 🤗

What it solves

Trackio is a lightweight, free experiment tracking library designed to replace heavy cloud-dependent tools. It specifically addresses the needs of both human researchers and AI agents by providing a local-first logging system that doesn't require account creation and offers high-throughput data storage for parallel experiments.

How it works

Trackio logs metrics, configurations, and media to a local SQLite database, which can be "frozen" into Parquet files for efficiency. It uses a non-blocking background thread to batch and write logs every 0.5 seconds, ensuring that training loops are not slowed down by I/O. For visualization, it provides a Gradio-inspired dashboard that can be run locally, self-hosted on private infrastructure, or deployed for free to Hugging Face Spaces for collaboration and embedding in websites via IFrames.

Who it’s for

  • ML Researchers: Those seeking a free, local-first alternative to Weights & Biases with a familiar API.
  • AI Agent Developers: Developers building autonomous ML experiments who need a programmatic CLI and SQL interface to query experiment data.
  • Collaborative Teams: Users who want to share live dashboards via Hugging Face Spaces without managing complex infrastructure.

Highlights

  • W&B Compatible: API compatible with wandb.init, wandb.log, and wandb.finish for easy drop-in replacement.
  • LLM-Friendly: Includes a CLI for programmatic access and direct SQL querying of experiment data.
  • Local-First Design: No account required to start logging; data stays local by default.
  • High Throughput: Non-blocking architecture allows thousands of log calls per second.
  • Extensible: Built with Python and Svelte 5, making it easy to fork and customize both the backend and the frontend dashboard.

Related

  • Dispatch
    Hugging Face Trackio Release

    Hugging Face has released Trackio, a free, open-source, lightweight experiment tracking library that serves as a drop-in replacement for wandb with native integration for Hugging Face Spaces and Datasets.

  • Dispatch
    Hugging Face Open-source DeepResearch

    Hugging Face has open-sourced a framework to reproduce OpenAI's Deep Research capabilities, achieving a 55.15% score on the GAIA benchmark using a code-native agentic approach.

  • Project
    huggingface/datasets

    A lightweight library for easy one-line loading and efficient pre-processing of massive multi-modal datasets for machine learning training and evaluation.

  • Project
    wandb/examples

    A curated set of ready‑to‑run scripts and Colab notebooks that show how to plug the Weights & Biases tracking library into popular ML frameworks (PyTorch, TensorFlow/Keras, Hugging Face, XGBoost, scikit‑learn, etc.). It demonstrates logging runs, configs, metrics, gradients, and artifacts, helping users add experiment tracking and reproducibility to their models with just a few lines of code.

  • Dispatch
    swift-huggingface release notes / what's new

    Hugging Face has released swift-huggingface, a comprehensive Swift client for the Hugging Face Hub that introduces reliable model downloads, a Python-compatible cache, and OAuth 2.0 support.