What an AI benchmark is: tasks, test sets, metrics and scoring rules
AI benchmarks are standardized tests that use predefined tasks, test sets, metrics, and scoring rules to evaluate machine learning models. They enable fair comparison, track progress, and enforce reproducibility. Examples include GLUE, MLPerf, and Dynabench.
Definition and components
An AI benchmark is a standardized evaluation framework that pairs tasks with fixed test sets, metrics, and scoring rules to compare model performance. The GLUE benchmark, for example, evaluates models across a diverse range of natural language understanding tasks and includes a hand-crafted diagnostic test suite, as described in the GLUE benchmark paper.
Benchmarks are designed to be model-agnostic, meaning any model can be tested under the same conditions. This generality allows results to be compared meaningfully across different approaches.
Metrics and test sets
Metrics quantify how well a model performs on a task. Accuracy, for instance, is the number of correct predictions divided by the total number of predictions, as defined in the Google Machine Learning Glossary.
Test sets are fixed collections of examples that are not used during training. They provide an independent measure of how well a model generalizes to unseen data.
Scoring rules and governance
Scoring rules define how results are reported and ensure that comparisons are fair. MLCommons, which oversees the MLPerf suite, sets rules on what model changes are allowed and requires submitters to join a working group, sign a contributor license agreement, and follow submission requirements, as stated on the MLCommons benchmark page.
Reproducibility is a core goal. Working groups define the exact model, dataset, and rules to ensure that results can be verified by others.
Dynamic and evolving benchmarks
Static benchmarks can become saturated, so some efforts create dynamic benchmarks. Dynabench is an open-source platform that enables human-and-model-in-the-loop dataset creation, where annotators try to create examples that a target model misclassifies but humans would not, as explained in the Dynabench paper.
Large-scale benchmark projects like BIG-bench aim to quantify and extrapolate the capabilities of language models across many tasks, as described in the BIG-bench paper.
| Benchmark suite | What it measures |
|---|---|
| AILuminate | Safety of general chatbot AI systems |
| AlgoPerf | Speed of training algorithms to a target performance |
| MLPerf Client | Performance of large language models on personal computers |
| MLPerf Inference Datacenter | Speed of processing inputs with a trained model |
| MLPerf Training | Speed of training to a target quality metric |
| MLPerf Storage | Speed of storage systems supplying training data |
Frequently asked questions
What is an AI benchmark?
An AI benchmark is a standardized evaluation that uses tasks, test sets, metrics, and scoring rules to measure model performance.
How are metrics chosen?
Metrics depend on the task; accuracy is common, but precision and recall are often used for imbalanced data.
What is a test set?
A test set is a fixed collection of examples not used during training, providing an independent evaluation.
Who creates benchmarks?
Research groups and organizations like MLCommons and academic teams create benchmarks.
Why do benchmarks need scoring rules?
Scoring rules ensure fair comparison and reproducibility, so results are trustworthy.