Fine-Tuning · Monitoring jobs
Docs / Fine-Tuning

Monitoring jobs

Watch a job's status and logs from the Jobs list and its detail view, or cancel and delete jobs.

Jobs train asynchronously on provisioned GPUs. The Jobs list shows every job on your account with its current status, and opening a job takes you to its detail view with live training logs.

Queued
waiting for a GPU
Training
live logs streaming
Succeeded
or failed / cancelled
Deploy or delete
your call

The Jobs list

Open Fine-Tuning → Jobs to see every job you've started, across every base model and method, along with its current status — queued, training, succeeded, failed, or cancelled.

A job's detail view

Click into a job to see its live status and streaming training logs update in real time as it runs, without needing to refresh the page.

Reading the training log

The log streams a line per training step, including the current loss — the number the training process is trying to push down. You don't need to watch it closely, but a quick glance tells you whether a job is on track:

  • Healthy — loss drops fairly steadily over the first portion of the run, then levels off into a gentle downward or flat trend as training continues. Small step-to-step noise is normal.
  • Worth a look — loss stays flat from the very first steps and never moves, which usually points at a learning rate that's too low or a dataset that isn't teaching the model anything new. Consider cancelling and adjusting hyperparameters rather than waiting out the full run.
  • Worth a look — loss spikes sharply upward partway through, or turns into NaN/inf. This typically means the learning rate is too high for the method chosen. Cancel the job and lower the learning rate before restarting.
Note
The exact loss curve depends heavily on your dataset and method, so treat these as general shapes to watch for, not a strict pass/fail threshold. When in doubt, let a short job run to completion and check the deployed model's behavior directly in the Playground.

Cancel or delete a job

From the job's detail view, click Cancel to stop a job that is still running and release its GPU. Once a job is in a terminal state, click Delete to remove it from your Jobs list.

Note
Cancelling still charges for the GPU time used up to that point — see Pricing.