Introducing Trackio: A Lightweight Experiment Tracking Library from Hugging Face

Hugging Face has introduced Trackio, an open‑source Python library that tracks machine‑learning experiments and visualizes metrics using a local Gradio dashboard. Designed as a drop‑in replacement for wandb, Trackio supports the same API calls—wandb.init, wandb.log, and wandb.finish—allowing users to import it as wandb in existing code. After logging experiments, users can launch a local dashboard and optionally sync it to a Hugging Face Space by providing a space_id during initialization. The synced dashboard can be shared publicly or privately via a URL or embedded with an iframe. Data logged to a Space is stored in an SQLite database that is reset on restart; Trackio automatically converts this database to a Parquet dataset and backs it up to a Hugging Face Dataset every five minutes, with an optional dataset_id parameter. The library also captures GPU energy usage by querying nvidia‑smi, enabling researchers to quantify and share energy consumption. Trackio integrates natively with Hugging Face libraries such as transformers and accelerate, and is currently in beta, lacking features like artifact management and complex visualizations. The Hugging Face science team reports that Trackio’s lightweight design facilitates rapid experimentation and easy sharing of training progress, making it a free alternative to more complex or paid tracking solutions.