Ground Truth
The objective reality of what the right answer is—what you measure model predictions against.
Think of it like
Like knowing the actual score of a game when you're evaluating a sports commentator's calls.
Example
For sentiment classification, ground truth is 'this review is angry' as labeled by trained annotators.
How it actually works
Ground truth is often not truly ground—it's human judgment, which is fallible. In noisy or ambiguous domains (does this summary capture the essence?), there's no perfect ground truth. Best practice is collecting multiple annotations per example and measuring inter-annotator agreement.
For product teams
Be explicit about what ground truth means for your task and whether it's expert judgment or crowd consensus.
For engineers
Store ground truth separately from predictions; implement versioning and audit trails.
Related
- Gold Set — Where it comes from.
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