api: YAML JSON TOON Database version: 1.0.0 format: yaml dataset: id: 2838 slug: llm-glossary title: Large Language Model Glossary description: "Key terms and concepts in LLM technology: tokens, context windows, temperature, RAG, fine-tuning, quantization, KV cache, and emergence." summary: "A glossary of large-language-model terms with precise definitions: tokens, context windows, temperature, retrieval-augmented generation, fine-tuning, quantization, KV cache and emergence, extended with embeddings, prompts, hallucination, sampling methods, RLHF, tool use, agents, mixture of experts, distillation, LoRA, perplexity and prompt injection." use_when: - You need a one-sentence definition of an LLM term for documentation or a prompt. - You are aligning a team on shared vocabulary. caveats: - The field changes fast; definitions here are conceptual and avoid claims about specific models. - Some terms (agent, emergence) are used loosely and with different meanings across papers and products. - Temperature zero does not guarantee reproducibility across hardware, serving systems, or model revisions. related: - attention-mechanisms - transfer-learning - ml-evaluation-metrics - ai-ethics-interpretability - llm-evaluation-failure-modes sources: - title: Stanford HELM evaluation framework url: "https://crfm.stanford.edu/helm/latest/" - title: Hugging Face generation parameters url: "https://huggingface.co/docs/transformers/main/en/main_classes/text_generation" - title: Schaeffer et al. on emergence and metric choice url: "https://arxiv.org/abs/2304.15004" reading: - llm-evaluation-checklist-for-reference-grounded-answers reviewed_at: "2026-10-12" provenance: Original editorial reference notes by YJTOON, informed by the linked primary documentation. External specifications retain their own licenses. entry_count: 41 fields: - section: terms field: term type: string example: Token - section: terms field: definition type: string example: A unit of text processed by LLMs, typically a w… category: AI & ML Terms category_slug: ai-ml-terms tags: llm,ai,glossary,nlp view_count: 65 created_at: 1777673262 updated_at: 1791799373 data: terms: - term: Token definition: A unit of text processed by LLMs, typically a word fragment or punctuation - term: Context Window definition: Maximum number of tokens an LLM can process in a single inference - term: Temperature definition: A parameter that rescales sampling logits; lower values concentrate probability. Zero-temperature behavior and reproducibility depend on the implementation. - term: RAG definition: "Retrieval-Augmented Generation: combining LLMs with external knowledge retrieval" - term: Fine-tuning definition: Adapting a pre-trained model to specific tasks with additional training - term: Quantization definition: Reducing model precision (e.g., FP16 to INT8) to decrease size and increase speed - term: KV Cache definition: Key-Value cache storing attention computations for efficient autoregressive generation - term: Emergence definition: A capability whose measured performance appears to change sharply with scale; the appearance can depend on the metric and evaluation design. - term: Embedding definition: A dense vector representing text or another input so that semantically similar items are close in vector space - term: Prompt definition: The input text, including instructions and examples, given to a model to elicit a response - term: System Prompt definition: Instructions set by the application developer that frame the model's role and behavior for a conversation - term: Hallucination definition: Output that is fluent and confident but false or unsupported by the provided sources - term: Top-p (Nucleus Sampling) definition: Sampling from the smallest set of tokens whose cumulative probability exceeds p, which adapts the candidate set to the distribution - term: Top-k Sampling definition: Sampling from only the k most probable next tokens - term: Inference definition: Running a trained model to produce outputs, as opposed to training it - term: Pretraining definition: The initial large-scale training of a model on broad data to learn general patterns before specialization - term: Instruction Tuning definition: Fine-tuning a model on examples of instructions and desired responses so it follows directions - term: RLHF definition: "Reinforcement learning from human feedback: tuning a model using human preference judgments as a reward signal" - term: Chain-of-Thought definition: Prompting or training a model to produce intermediate reasoning steps before a final answer - term: Few-Shot Prompting definition: Including a handful of worked examples in the prompt to show the desired task format - term: Zero-Shot definition: Performing a task from instructions alone, with no examples in the prompt - term: In-Context Learning definition: A model's ability to adapt to a task from examples or instructions in the prompt without updating its weights - term: Tool Use (Function Calling) definition: A model emitting structured calls to external functions or APIs, whose results are returned to it - term: Agent definition: A system in which a model repeatedly decides on actions, uses tools and observes results to pursue a goal - term: Mixture of Experts definition: An architecture that routes each token to a few specialized sub-networks, increasing capacity without activating all parameters - term: Distillation definition: Training a smaller model to imitate the outputs of a larger one - term: LoRA definition: "Low-rank adaptation: fine-tuning by training small added matrices while keeping the base weights frozen" - term: Perplexity definition: A measure of how well a language model predicts text; lower means the text is less surprising to the model - term: Prompt Injection definition: An attack that places instructions in untrusted input so the model follows them in place of the developer's - term: Guardrails definition: Checks and constraints around a model that filter inputs and outputs to keep behavior within policy - term: Multimodal definition: Able to process more than one kind of input or output, such as text, images and audio - term: Vector Database definition: A store optimized for nearest-neighbor search over embeddings - term: Chunking definition: Splitting documents into pieces of suitable size for embedding and retrieval - term: Logits definition: The raw, unnormalized scores a model produces for each possible next token before softmax - term: Alignment definition: The effort to make model behavior match human intentions and values - term: Scaling Laws definition: Empirical relationships showing how model loss improves predictably with more parameters, data and compute - term: Model Context Protocol (MCP) definition: An open protocol that standardizes how applications give models access to tools and data sources - term: Groundedness definition: Whether answer claims are supported by the supplied evidence; it differs from whether the evidence itself is correct. - term: Calibration definition: Agreement between stated or estimated confidence and observed correctness over a set of predictions. - term: Held-out Evaluation definition: Assessment on cases withheld from development choices, to reduce optimistic bias from repeated tuning. - term: Abstention definition: Declining to answer when the evidence or confidence is insufficient, according to an explicit policy.