ToolJoltTools

Cohen's Kappa Calculator

Inter-rater / model-vs-label agreement corrected for chance — the metric for labeling quality and annotation studies.

Observed agreement (%)
Chance agreement (%)
Cohen's κ

Landis & Koch bands: <0.20 slight, 0.21–0.40 fair, 0.41–0.60 moderate, 0.61–0.80 substantial, 0.81–1.00 almost perfect. Raw agreement % alone is misleading when one class dominates — κ corrects for it.

Formula

κ = (pₒ − pₑ) / (1 − pₑ) — observed agreement minus chance agreement, normalized by the room above chance
References: Cohen (1960), A Coefficient of Agreement for Nominal Scales; Landis & Koch (1977), The Measurement of Observer Agreement for Categorical Data

About Cohen's Kappa Calculator

When two annotators (or a model and a gold label) agree 90% of the time, is that good? Not if the easy majority class is 88% of cases — they'd hit 88% by both guessing it blindly. Cohen's kappa subtracts that chance agreement and rescales, so κ measures agreement BEYOND luck. It's the standard metric for labeling-quality audits, inter-annotator reliability and dataset validation. Enter the 2×2 agreement table and this calculator returns observed agreement, chance agreement and κ with the Landis-Koch interpretation band.

How to use Cohen's Kappa Calculator

  1. 1Enter your values into Cohen's Kappa Calculator — sensible, domain-typical defaults are pre-filled so you see a real result immediately.
  2. 2The result recomputes live using the formula shown on the page; there is no button to press.
  3. 3Adjust any input to compare scenarios, then read the worked example to see the substituted numbers.

Why use Cohen's Kappa Calculator?

  • Computes Cohen's Kappa instantly in your browser — no sign-up, no upload, no server round-trip.
  • 100% free and unlimited, with the exact formula shown: κ = (pₒ − pₑ) / (1 − pₑ) — observed agreement minus chance agreement, normalized by the room above chance.
  • Runs entirely client-side, so every value you enter stays private on your device.
  • Live recompute as you type, with a worked example and authoritative references for trust.

Frequently asked questions

Why not just report raw agreement percentage?+

Because chance inflates it. On a task that's 95% one class, two raters who never look will agree ~90% of the time. Kappa removes that floor — κ=0 means agreement is exactly what chance predicts, κ=1 means perfect agreement. It's the honest measure of whether your guidelines actually align raters.

What kappa is 'good enough' for a labeled dataset?+

Convention treats κ ≥ 0.8 as reliable enough to trust labels, 0.6–0.8 as usable with caution, and below 0.6 as a sign your annotation guidelines are ambiguous and need revision before scaling labeling. The exact bar depends on task difficulty and stakes.

What is the 'kappa paradox'?+

On highly imbalanced data, κ can be low even when observed agreement is very high, because there's little room above chance. This is real, not a flaw — it's telling you the agreement isn't impressive given the imbalance. For such cases, report κ alongside prevalence, or consider Gwet's AC1.

How does this extend to more than two raters or categories?+

For multiple raters use Fleiss' kappa; for ordinal categories use weighted kappa (penalizing distant disagreements less). This calculator handles the foundational two-rater, two-category case — the building block for understanding all of them.

Related tools

Related ML & AI tools

Sponsored