📖 Some key terms around LLMs from papers & other resources discovered on my LLM journey, re-written as one-liners.
⭐ Large Language Model (LLM) is a deep learning model like ChatGPT, based on transformer architecture, trained on huge amounts of text
⭐ Explicit prompts are clear instructions including the role, task, & output format, to guide an LLM’s behaviour
⭐ Implicit prompts guide an LLM’s behaviour without explicit instructions, relying on model general understanding, for more creativity
⭐ Prompt engineering crafts prompts to guide the LLM for best possible output
⭐ Completion is an LLM’s output in response to a prompt
⭐ Generative AI is an AI instructed by a prompt input to create original content from existing data.
⭐ Hallucination is when an LLM generates outputs that sound plausible but incorrect
⭐ Chain-of-thought prompting aims to solve complex problems by decomposing prompts into intermediate step prompts
⭐ One-shot learning for a model trained to understand new concepts with only a single example
⭐ Few-shot learning for a model trained to understand new concepts with only a few examples
⭐ Foundational model is a pre-trained model like GPT-3 serving as a starting point for downstream tasks like information retrieval
⭐ Plugins / agents enable LLMs to access APIs for powerful capabilities like performing web searches for an up-to-date world view or fact-checking
⭐ Retrieval Augmented Generation (RAG) is the process of supplementing a prompt with additional information based on web searches or queries of internal documents
⭐ Vector database is a specialized type of dB designed to store & efficiently retrieve vector data based on similarity of, for e.g., an output from an LLM