Out-of-distribution
Input examples that don't match the distribution your model saw during training.
Think of it like
Showing someone a cat photo when they've only ever seen dogs.
Example
Your model is trained on English news. A user asks in Spanish. The Spanish input is out-of-distribution.
How it actually works
OOD inputs cause poor model behavior because the model is extrapolating past its training data. Some models are robust (they say "I don't know"), others hallucinate confidently. Detecting OOD without ground truth is hard — you can't just check if the input looks weird because the model's embedding space might not preserve "weirdness" in a useful way.
For product teams
Plan for OOD gracefully. Don't let users hit it without warning.
For engineers
Use uncertainty estimates, entropy, or embedding distance to detect OOD. Fallback to a simpler system.
Related
- Opposite of in-distribution.
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