Glossary term
Glossary term
Evaluation and Benchmarks
An all-or-nothing metric in which the model's output either matches ground truth or the reference text exactly or it doesn't. For example, if ground truth is orange, the only model output that satisfies exact match is orange.
Exact match can also evaluate models whose output is a sequence (a ranked list of items). In general, exact match requires the generated ranked list to exactly match ground truth; that is, each item in both lists must be in the same order. That said, if ground truth consists of multiple correct sequences, then exact match only requires model's output matches one of the correct sequences.
Created for this library
A question-answering team reports exact match on its evaluation set so model owners see a strict measure of correctness on factual questions.
A search team uses exact match as one of several metrics for short factoid queries where rewording is unhelpful.
A code-completion team reports exact match between the model's suggestion and the developer's accepted edit on a fixed benchmark.
Definition source: Google for Developers Machine Learning Glossary | Creative Commons Attribution 4.0 License