Large language model

A large language model is a neural network trained on extensive text data to predict tokens and thereby process or generate language.

Transformer: A neural-network architecture that uses attention to process sequences in parallel. Most modern large language models use transformer layers to connect information across a text sequence.
Language model: A model that assigns probabilities to sequences of linguistic units. Large language models belong to this broader family of models.
Machine translation: The automated conversion of text or speech from one language to another. Language models can translate by predicting text in a target language from source-language context.
Hallucination (artificial intelligence): An AI system's production of plausible-sounding but unsupported or false information. Fluent generation does not guarantee that a model's factual claims are correct.
Algorithmic bias: Systematic, unjust differences in an algorithm's outputs or effects across groups. Patterns and stereotypes in training data can shape unequal model responses.
Self-attention: A mechanism that lets each sequence element weight information from other elements. It lets a model use relevant words elsewhere in the input when processing each token.
Neural network: A computational model made of connected units with adjustable parameters. Large language models are neural networks trained to model language.
Text summarization: The production of a shorter text that preserves key information from a source. A model can condense long passages into summaries, though important details may be lost.
Knowledge graph: A structured representation of entities and their relationships. Unlike a language model's distributed parameters, a knowledge graph makes relations explicit.
Data privacy: The protection and appropriate handling of information about identifiable people. Training datasets may contain personal information, raising questions about consent and exposure.

Linked from 10 pages