Rollouts

The Rollouts page lets you inspect every sample the model generated during training — the prompts, completions, rewards, and per-sample metrics. There is also a separate Discarded Rollouts page that shows samples that were thrown away (e.g. due to async staleness or zero advantage), with a badge indicating the discard reason.

​

Browsing samples

The left sidebar lets you navigate through samples by training step. Within each step, samples are organized into groups — each group shares the same prompt, and you can click into a group to see all the individual completions and how they compare. Metric columns in the sidebar show color-coded dots for each sample’s reward, advantage, token count, and any custom metrics, so you can quickly scan for patterns across a group.

​

Sample view

Selecting a sample shows the full rollout in the main panel:

• Prompt — system prompt and user prompt with token counts
• Completions — the model’s response, with multi-turn conversations shown as separate turns. For agentic environments, you’ll see the alternating assistant/user (tool) turns
• Rewards — total reward and per-metric breakdowns, color-coded on a red-to-green scale based on the environment’s reward range
• Advantage — the sample’s advantage within its group
• Golden answers — reference answers when available, so you can compare against the expected output

The Render menu lets you toggle formatting for think blocks, markdown, LaTeX, and code blocks — useful for inspecting reasoning traces or structured outputs. You can also view the raw text of the full sample as a single string.

​

Metrics sidebar

The right sidebar shows training metric charts alongside the rollouts, so you can correlate what the model is generating with how metrics are evolving. You can click on a chart to jump to that step’s rollouts. It supports the same metric catalog, EMA smoothing, and multi-run overlay as the Metrics page. Charts in the sidebar can be reordered with drag-and-drop — hover over a chart to reveal the drag handle. You can also add the same plot multiple times with independent filter settings, useful for comparing different views of the same metric (e.g., with and without outlier filtering). Each chart in the sidebar has per-chart controls matching the Metrics page: ignore outliers, ignore first step, and manual Y-axis range.