Skip to main content
Categorical {bars} rollups let you compare many related scalars — per-class accuracy, per-shard throughput, per-expert load — as a single distribution instead of dozens of overlapping line charts.

​
Producing bar charts

Log three or more sibling scalars under a common prefix and Pluto automatically offers that prefix as a {bars} rollup — no special SDK call needed:
Pluto detects the eval/per_class prefix and surfaces it as eval/per_class/{bars}. If you control the SDK and want true binned distributions, pluto.Histogram is the richer option.

​
Viewing

Bar charts live in the Distributions widget alongside numeric histograms. To add one, open Add Widget, choose the Distributions tab, and search for your <prefix>/{bars} group. Adding a Distributions widget from the Add Widget modal

​
View modes

Each widget renders in one of three modes. Step — a bar chart at the scrubbed step, bars sorted by global max value. Use the step slider below the widget to scrub through training. Bar chart in Step mode — one bar per category at a single step Ridgeline — TensorBoard-style joyplots, one ridge per step, so you can watch each category’s value evolve over training. Bar chart in Ridgeline mode — one ridge per step Heatmap — a density grid: value as color, category on the X axis and step on the Y axis. Bar chart in Heatmap mode — value as color A Y: step / Y: run toggle switches whether ridges and heatmap rows are stacked by step or by run.

​
Steps on X

In Ridgeline and Heatmap modes, the Steps on X toggle transposes the chart so training steps run along the X axis and categories stack down the Y axis. A {bars} heatmap aligns its X range with the line chart above it, so a metric and its distribution read as one coordinated view across time. Bar-chart ridgeline with Steps on X — steps run along the X axis The same transpose applies in Heatmap mode: Bar-chart heatmap with Steps on X

​
Settings

The settings popover has an Ignore outliers toggle (on by default) that clamps the shared frequency scale so one extreme step doesn’t flatten the rest into a baseline. It’s inert in Step mode, where the axis auto-scales to the current snapshot.