Evaluation icons

The 16 evaluation icons in AI & LLM, described

Ablationablation

An ablation — take one part of a system out and see what breaks.

Accuracyaccuracy

Accuracy — how often the model gets it right, hits on target.

Benchmarkbenchmark

A benchmark — how a model scores against others on a shared test or leaderboard.

Best of Nbest-of-n

Best of N — generate several candidates and keep the winner as the answer.

Confusion matrixconfusion-matrix

A confusion matrix — what a classifier got right and which classes it mixed up.

Degradationdegradation

Degradation — performance getting worse over time, a decline or a decay.

Driftdrift

Drift — a model and the world slowly coming apart as the data it sees changes.

Eval suiteeval-suite

An eval suite — a whole battery of checks run together as one harness.

Evaluationevaluation

Evaluation — measuring how good a model's output is with tests, scores and benchmarks.

F1 scoref1

F1 score — precision and recall forced to agree in one harmonic number.

Golden setgolden-set

A golden set — the reference answers you trust enough to grade against.

Human evaluationhuman-eval

Human evaluation — a person rating and scoring the model's answers.

Model comparisonmodel-comparison

Model comparison — this model against that one, benchmarked side by side.

Perplexityperplexity

Perplexity — how surprised a model is by the text, a measure of uncertainty.

Precisionprecision

Precision — how many of the hits were actually right, tight and accurate.

Reward modelreward-model

A reward model — the model that scores another model's answers as a judge.

In code, each is one import — import { Ablation } from "@iconmind/react/icons/ablation" — and the same name in Vue, Svelte, Solid, Preact, React Native, Astro, Blade and Flutter.

Narrow it by tag