Search Results for "inference"

RETRO (Retrieval-Enhanced Transformer) is a large language model architecture developed by OpenAI that augments transformer models with a retrieval mechanism. Instead of relying solely on learned parameters, RETRO retrieves relevant documents from a large external database during inference, allowing it to ground responses in external knowledge. This design improves factual accuracy, reduces hallucinations, and enables smaller models to perform comparably to much larger ones by leveraging retrieval. The repository provides code and resources for training and evaluating RETRO models, along with infrastructure for integrating retrieval into the transformer pipeline. ...