What term describes an n-gram that the LLM wasn't trained on?

The term that describes an n-gram the language model wasn't trained on is "Out-of-Vocabulary" (OOV). This term refers to words or sequences of words that do not appear in the training data of the model. In the context of natural language processing, vocabulary consists of the set of words that the model recognizes and can generate. When the model encounters an n-gram made up of tokens that were not part of its training corpus, it will consider this n-gram as out-of-vocabulary.

Being OOV can pose challenges for language models, as these models are typically designed to predict the next token based on the preceding tokens they have learned from their training. Encountering OOV n-grams can lead to difficulties in generating coherent and contextually appropriate text, since the model lacks familiarity with those specific combinations of words.

The other terms do not accurately describe the situation. "In-Sentence Construct" would refer to words or phrases that are used in a sentence but does not denote whether they were trained on or not. "Language Model Gap" might imply limitations or areas where the language model performs poorly, but it does not specifically refer to unrecognized n-grams. "Contextual Drift" indicates a change in understanding or relevance