Trackio: A Lightweight Experiment Tracking Library from Hugging Face (huggingface.co)

Hugging Face has released Trackio, a free, open-source, lightweight Python experiment-tracking library that acts as a drop-in replacement for tools like wandb. Trackio provides a local Gradio dashboard (launchable with trackio.show or from Python), seamless API compatibility with wandb.init / wandb.log / wandb.finish (you can simply import trackio as wandb), and one-click syncing to Hugging Face Spaces for easy sharing or embedding (iframe support). It integrates natively with transformers.Trainer and accelerate, logs GPU metrics (via nvidia-smi), and backs up Space-session SQLite logs to Parquet datasets on Hugging Face every five minutes to avoid data loss when a Space restarts. For the AI/ML community this matters because Trackio is local-first, transparent, and avoids vendor lock-in: data is easy to extract for custom analysis, energy usage can be quantified and added to model cards, and the core codebase is intentionally small (under 1,000 lines) so teams can extend it. Technical trade-offs: it’s beta and lacks advanced features like artifact management and complex visualizations found in larger platforms, but its tight integration with Hugging Face Datasets/Spaces, compatibility with existing training APIs, embeddability, and free hosting make it a pragmatic choice for reproducible, shareable, and low-friction experiment tracking.