All view
The default view organizes every available metric into sections:- Custom Metrics — metrics logged by the trainer with section/group structure (keys use
/delimiters likesection/group/metric) - Reward — reward sum stats (mean, std, min, max)
- Samples Metrics — per-reward-function breakdowns from your environment’s
sample_metrics - Advantage — advantage statistics with a zero-line reference
- Evals — one subsection per eval, showing eval-specific metrics
- Inference Performance — time-bucketed bar charts showing throughput metrics: inference calls/min, requests done/min, rollout groups done/min (total, kept, discarded, canceled). Configurable bucket interval and optional step completion lines
- Rollouts — token length distributions for prompts, completions, and totals
- Discarded Rollouts — discard counts, zero-advantage breakdowns, and token metrics for discarded samples
- Timeline — step timing breakdowns (forward, backward, loss, KL, weight sync, etc.) at both full-step and per-microbatch granularity
Custom view
A fully editable dashboard where you can build your own layout. Create sections and groups, add plots from the metric catalog, and reorder everything with drag-and-drop. You can add the same plot multiple times with independent filter settings — useful for comparing different views of the same metric. The layout is saved server-side, so it persists across sessions and devices.Per-chart controls
Each chart has its own settings accessible via the settings icon:- Ignore Outliers — hides extreme values using IQR-based filtering so you can focus on the meaningful range
- Ignore First Step — excludes step 0, which is often much slower due to compilation. Enabled by default for timing metrics
- Min Y / Max Y — manually set the y-axis range to zoom into a specific value range