Gradient Comparison Table

Tabular comparison of analytical vs numerical gradients across operations — used to validate autograd implementations.

Preview

Gradient comparison: 0 of 4 rows agree within tolerance. x + y: analytical 1.000000, numerical 1.000010, fail. x · y: analytical 0.500000, numerical 0.499980, fail. x²: analytical 4.000000, numerical 4.000020, fail. softmax: analytical 0.210000, numerical 0.180000, fail.
Op	Analytical	Numerical	Error	Match
x + y	1.000000	1.000010	1.000e-5	Fail: exceeds tolerancefail
x · y	0.500000	0.499980	2.000e-5	Fail: exceeds tolerancefail
x²	4.000000	4.000020	2.000e-5	Fail: exceeds tolerancefail
softmax	0.210000	0.180000	0.030000	Fail: exceeds tolerancefail

Anatomy

Each row carries an op label, the analytical gradient, the numerical estimate, and an optional per-row tolerance. Rows whose absolute error exceeds the tolerance flip to cb-error and the badge changes from "match" to "fail".

Props

Prop	Type	Default	Description
`rows`	`GradientComparisonRow[]`	required	Comparison rows.
`showError`	`boolean`	`true`	Render the error column.

Accessibility

<table> with a caption summarising the overall pass/fail. Each row exposes data-cb-status for open styling.