Training runPro
Follow a fine-tune from queue to endpoint: loss for training and held-out data is plotted against the full step range (which also shows epoch progress), alongside learning rate, time spent and time left, the latest log lines, Pause, a Cancel that asks twice, and Deploy at the end.
oxyn add training-runsupport-replies-v4
Qwen3 14B · LoRA r16 · 8× H100
Waiting for 8× H100 · 2nd in queue
- Learning rate
- 3.33e-6
- Elapsed
- 0m 00s
- ETA
- About 46m 00s
- Train loss
- —
The preview stays free for everyone. Pro opens the code of every Pro component, screen, agent-app template, block and kit, plus the Pro CLI.
Rendered from Rust at build time · goes live when visible
When to use
- Follow a fine-tuning job from the queue to a deployed endpoint, with the loss curve as the main signal.
When not to use
- Replays a multi-agent run on one clock: use Agent timeline instead.
- One model at a glance: use Model card instead.
- A small automation on a single rail: use AI workflow instead.
Installation
Training run is part of Oxyn Pro; the preview is free. Pro items install with the oxyn CLI once your licence is signed in; until then Pro members copy the code from this page.
oxyn add training-runUsage
The preview is free. Pro opens the code of every Pro component, screen and template.
Examples
Training
A LoRA run from the queue to Deploy.
support-replies-v4
Qwen3 14B · LoRA r16 · 8× H100
Waiting for 8× H100 · 2nd in queue
- Learning rate
- 3.33e-6
- Elapsed
- 0m 00s
- ETA
- About 46m 00s
- Train loss
- —
Code Pro
The preview is free. Pro opens the code of every Pro component, screen and template.
Oom
Fails at step 940; Retry from checkpoint 800.
support-replies-v4
Qwen3 14B · LoRA r16 · 8× H100
Waiting for 8× H100 · 2nd in queue
- Learning rate
- 3.33e-6
- Elapsed
- 0m 00s
- ETA
- About 46m 00s
- Train loss
- —
Code Pro
The preview is free. Pro opens the code of every Pro component, screen and template.
Quiet
No event log, quarter speed (Oxyn extra).
support-replies-v4
Qwen3 14B · LoRA r16 · 8× H100
Waiting for 8× H100 · 2nd in queue
- Learning rate
- 3.33e-6
- Elapsed
- 0m 00s
- ETA
- About 46m 00s
- Train loss
- —
Code Pro
The preview is free. Pro opens the code of every Pro component, screen and template.
API reference
TrainingRun
| Prop | Type | Default | Description |
|---|---|---|---|
nowrequired | ClockMs | — | Host clock. |
oom | bool | Default::default() | Fails out of memory at step 940 (Retry demo). |
speed | u64 | 100 | Playback speed, percent. |
events | bool | true | Show the event log. |
Read from the Rust signature and doc comments.
Keyboard interactions
| Key | Action |
|---|---|
| Tab | The loss chart, then the footer actions |
| ←or→ | Step through samples on the chart (Shift jumps ten) |
| HomeorEnd | First or latest sample |
| Escape | Back to the latest sample |
| EnterorSpace | Press an action; Cancel needs Confirm cancel |
Accessibility
- Status changes are announced (Queued, Training, Paused, Completed, Failed), never every step.
- The job state is shown as a glyph with a label; why it failed is written out at the bottom and in the log.
- The plot is a named group that speaks the step under the guide and both loss values; the figures beneath are a dl.
From the spec
Status changes are announced, never every step; the chart is a labelled group whose name carries the inspected step; stats are a description list; Cancel needs a second press. Every action returns the line the polite live region announces. The focus outline shows only for keyboard focus.
Motion & interaction
The smoothed train loss draws over its faint raw samples with validation dashed; epoch bands under the plot fill as progress. Events slide in newest first. Arrow keys inspect a sample with a guide line.
Timing
Everything on screen is derived from the host clock now (no timers in the component), so playback, tests and the GPUI view agree.
| Transition | Duration | Easing |
|---|---|---|
| Queue | 1400 ms | — |
| Start-up | 900 ms | — |
| Step | 18 ms per step at 1× | linear |
| Deploy | 1400 ms | — |
| Cancel confirm window | 3000 ms | — |
Reduced motion: loops and travel stop; each state shows its end pose and its words.
Responsive behaviour
- Grows to 640 px; on narrow cards the figures form a 2 × 2 grid and the key abbreviates to Val.
- The axis spans the whole run and the y range is fixed, so nothing rescales mid-run.
Edge cases (spec)
The x axis spans the whole run and the y range is fixed, so nothing rescales. Under 440 px (container query) stats go two by two and the legend shortens to Val; the footer wraps under the dataset. Checked at 1440, 768 and 390 px in light and dark with no text overlap. Chrome is monochrome grey; colour is semantic only (success, warning, danger, info) plus the official coloured brand and model marks.
Notes for AI
- Import
TrainingRunfromoxyn_dioxusand includeSTATUS_INDICATOR_CSSplus the token stylesheet once. - A GPUI version exists:
AiKitViewinoxyn_gpui. - Props to set first:
now. - It reads a clock (
now) for live timers; pass one clock for the whole page rather than one per component. - Keep its states and timings as they ship; restyle it through the CSS variables rather than by hand.