Job detail

The eight tabs on a running or finished job, and which one answers which question.

Open any job from the job list. Eight tabs, each answering a different question.

Charts: reward, accuracy and training scalars, step by step.

The tabs

Tab	The question it answers
Charts	Is it learning? Loss, reward, KL, gradient norm, entropy
Validation	What is it actually producing? Sample conversations and reward histograms
System	Is the hardware healthy? GPU utilisation, memory, network, per node
Config	What exactly was submitted? Resolved config and hyperparameters
Logs	What is it saying? The full stream, searchable and downloadable
Environment	What was it running in? Package versions, CUDA, GPU models, image digest
Artifacts	What did it produce? Checkpoints, exports, evaluation reports
Diagnosis	Why does this look wrong? An automated analysis, where enabled

Charts

Which curves appear is not a console decision — the recipe declares them in its metric contract, and the first two go on the overview chart. That is why two methods show different defaults and why a framework rename does not silently blank a chart.

Empty charts with moving logs means reporting is not happening, not that training is broken. See reporting from training code.

Validation

Validation samples with rewards, and the reward distribution.

Sample conversations with their rewards, plus the distribution. Algorithm-specific fields — a DPO rejected completion, an SFT reference answer — appear alongside without the platform needing to know what they are.

Reading the histogram is often faster than reading the reward curve: a mean that is climbing while the distribution splits into two clumps usually means the model found one easy subset and stopped improving on the rest.

Diagnosis

Diagnosis: findings with confidence, and a proposed follow-up.

Runs automatically on failure and on a schedule while training, if the deployment enables it. Each finding carries a confidence level, and a proposal may come with a submit plan — a concrete follow-up job you approve or reject.

Approving a plan submits a real training job and spends real quota. The dialog says so; read the diff before approving.

What you can do

Action	Effect
Pause	Stops the job, keeping its checkpoint. Resuming continues from it
Resume	Restarts from the latest checkpoint under the same run id
Stop	Ends the run. SIGTERM first, with a grace period to save a checkpoint
Share	Creates a read-only public link — see sharing
Diagnose	Runs an analysis now rather than waiting for the schedule