Compare training runs in Trackio with AI
Log the measurements from a training run, compare them in a dashboard and inspect alerts from a particular point in the experiment.
To compare training runs in Trackio with AI, give each run a recognisable name and log the same measurements as it progresses. The huggingface-trackio skill connects the Python logging calls with a dashboard and a command-line view of the results.
An engineer trying two learning rates might care less about the final loss than the point where one run became unstable. A warning recorded at that step gives the metrics some context. Local storage and a shared Hugging Face Space are both documented destinations.
What it does
- Record loss and accuracy at chosen training steps.
- Keep multiple named runs under one project.
- Look up a specific metric from the command line.
- Raise and inspect warnings about unusual values.
- Sync selected results to a Space of appropriate visibility.
Install and use
- Install the skill and Trackio in the training environment. For remote sync, connect an authorised Hub account and check the chosen Space’s visibility.
- Ask “Record loss by step for these two configurations and flag a sharp increase. Keep both runs in this project.”
- Read each run’s metrics and alerts together. Check the destination before enabling optional Slack or Discord notifications.
FAQ
Source
Made by Hugging Face. Original skill: huggingface-trackio, published on GitHub under the Apache-2.0 licence.
View on GitHub