Exact Match
A model's output is right if it exactly matches the expected answer, character for character.
Think of it like
Like a spelling bee where 'color' and 'colour' are different, even though they mean the same thing.
Example
On a reading comprehension benchmark, the model answers '2023' but the gold answer is '2023.' – exact match fails even though the substantive answer is correct.
How it actually works
Exact match is fast to compute and unambiguous, but brittle. It punishes minor variations and doesn't reward partially correct answers. It's fallen out of favor for tasks where multiple phrasings are valid. Remains useful for closed-answer tasks like named entity extraction or question answering where the set of correct answers is small and exact.
For product teams
Use exact match for well-defined closed-answer tasks, but switch to semantic similarity for open-ended questions.
For engineers
Implement string-level matching or token-level F1 as a baseline; graduate to semantic metrics as task complexity grows.
Related
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome