Out-of-Distribution Sample
An input sample whose distribution differs markedly from the data seen during training or compilation. It is a key concept for evaluating the generalization ability of a model.
Timeline (last 90 days)
Out-of-distribution samples refer to inputs whose distribution differs markedly from the data seen during training or compilation; local logic consisting of rules or small classifiers often fails near the boundaries.
Common engineering measures to mitigate the out-of-distribution sample problem include: setting confidence thresholds for compiled logic and falling back to the LLM at low confidence, continuously collecting online samples for regression evaluation, and maintaining a shadow mode in which the LLM and local logic run in parallel to monitor deviation.
All Facts (2)
Out-of-distribution samples refer to inputs whose distribution differs markedly from the data seen during training or compilation; local logic consisting of rules or small classifiers often fails near the boundaries.
50%UnverifiedCommon engineering measures to mitigate the out-of-distribution sample problem include: setting confidence thresholds for compiled logic and falling back to the LLM at low confidence, continuously collecting online samples for regression evaluation, and maintaining a shadow mode in which the LLM and local logic run in parallel to monitor deviation.
50%