Hugging Face has introduced Trackio, a new open-source Python library for experiment tracking designed to be lightweight, transparent, and easy to integrate. Built as a drop-in replacement for Weights & Biases (wandb), Trackio offers local dashboards by default and seamless syncing with Hugging Face Spaces for sharing and collaboration.
The library is under 1,000 lines of code, making it hackable and extensible. Logs persist locally in SQLite and are automatically backed up to Parquet datasets on Hugging Face every five minutes when synced. Trackio also integrates with Hugging Face libraries such as transformers and accelerate, enabling minimal-setup logging for training runs.
Key features include:
- API compatibility with wandb, enabling rapid migration.
- Local-first design: logs and dashboards run and persist locally by default, with the option to host on Hugging Face Spaces.
- Transparency: direct tracking of GPU energy usage via nvidia-smi, with results easily included in model cards.
Trackio emphasizes reproducibility and accessibility, providing researchers with a straightforward way to log and share experiments without relying on proprietary services.
Experiment tracking is a routine part of machine learning workflows, but the Hugging Face team argues that lowering barriers to entry is crucial for wider adoption and reproducibility.
Several researchers on the launch team pointed to the importance of transparency. They noted that Trackio’s ability to log GPU energy consumption and add it directly to model cards could set a baseline for reporting environmental impact across ML projects.
Some researchers raised questions about integration with the existing Hugging Face tooling. For instance, Ahmad Khan asked whether Trackio supports Nanotron. Tom Aarsen, a machine learning engineer at Hugging Face, replied:
Not yet, it looks like. I suspect it will be added in the future, let me ping the maintainers.
Others pointed out missing features compared to established tools, such as artifact management and advanced visualization. Hugging Face acknowledged these limitations, describing Trackio as a beta release meant to evolve through community contributions.
Trackio is now available on GitHub and PyPI, with Hugging Face encouraging feedback to guide development. By keeping the codebase small and the format open, the company hopes to foster a more transparent and flexible experiment tracking ecosystem. Users who want additional functionality are encouraged to submit feature requests through the project’s GitHub issue tracker.