AI & LLM Glossary

A neural network architecture that uses self-attention mechanisms to process sequential data. Introduced in the 'Attention is All You Need' paper (2017), it forms the basis of modern LLMs like GPT and BERT.

The process of taking a pre-trained model and training it further on a specific dataset to adapt it for a particular task or domain.

A parameter-efficient fine-tuning technique that freezes the original model weights and trains small adapter layers, reducing memory requirements by up to 90%.

The practice of designing and optimizing text prompts to get desired outputs from language models. Includes techniques like few-shot learning, chain-of-thought, and role-playing.

A technique that combines information retrieval with language generation. The model retrieves relevant documents from a knowledge base before generating a response.

The basic unit of text that a language model processes. Roughly 4 characters or 0.75 words in English. Models have maximum token limits (context windows).

A parameter (0-2) that controls randomness in model outputs. Lower values (0.1-0.5) make outputs more focused and deterministic, higher values (0.8-1.5) make them more creative.

When a language model generates false or nonsensical information that sounds plausible. A major challenge in deploying LLMs for factual tasks.

Dense vector representations of text that capture semantic meaning. Similar texts have similar embeddings, enabling semantic search and clustering.

The maximum number of tokens a model can process at once (input + output). Modern models range from 4K (GPT-3.5) to 2M (Gemini 1.5 Pro).

Providing a model with a few examples in the prompt to demonstrate the desired task, without fine-tuning.

A prompting technique where you ask the model to show its reasoning step-by-step, improving performance on complex tasks.

Reducing the precision of model weights (e.g., from 16-bit to 4-bit) to decrease memory usage and increase inference speed, with minimal accuracy loss.

A training method where human preferences are used to fine-tune models, making them more helpful, harmless, and honest.

The initial instruction that sets the model's behavior, role, and constraints. Not visible to end users but crucial for consistent outputs.

A technique that allows models to focus on different parts of the input when producing each part of the output. The foundation of transformer models.

A specialized database designed to store and efficiently search high-dimensional vectors (embeddings). Essential for RAG and semantic search applications.

The ability of a model to perform a task without any examples, using only the task description in the prompt.

A sampling method that considers the smallest set of tokens whose cumulative probability exceeds P. More dynamic than Top-K sampling.

A sampling method that restricts the model to choose from only the K most likely next tokens.

The process of breaking text into smaller units (tokens) that the model can process. Different models use different tokenization strategies.

A tokenization algorithm that iteratively merges the most frequent pairs of characters or character sequences. Used by GPT models.

An AI system that can perceive its environment, make decisions, and take actions autonomously. Often uses LLMs for reasoning and tool use.

The ability of LLMs to generate structured outputs that trigger external functions or APIs. Enables agents to interact with tools.

An architecture where multiple specialized sub-models (experts) are trained, and a gating mechanism decides which experts to use for each input.

Techniques that fine-tune only a small subset of model parameters, reducing memory and compute requirements.

Quantized LoRA - combines 4-bit quantization with LoRA to enable fine-tuning of large models on consumer hardware.

Fine-tuning a model on a dataset of instruction-response pairs to improve its ability to follow instructions.

The process of making AI models behave in ways that are helpful, harmless, and honest - aligned with human values and intentions.

An alignment approach where models are trained to follow a set of principles (a constitution) through self-critique and revision.

Search based on meaning rather than exact keyword matching. Uses embeddings to find semantically similar content.

A metric for measuring similarity between two vectors, commonly used to compare embeddings. Ranges from -1 (opposite) to 1 (identical).

A metric that measures how well a language model predicts text. Lower perplexity indicates better prediction capability.

An architecture with two components: an encoder that processes input and a decoder that generates output. Used in translation and summarization.

An architecture that only uses the decoder part of transformers. Most modern LLMs (GPT, LLaMA) are decoder-only models.

A generation method where the model predicts one token at a time, using previously generated tokens as context.

A search algorithm that keeps track of the top K most likely sequences at each step, balancing quality and diversity.

Always selecting the most probable next token. Fast but can lead to repetitive or suboptimal outputs.

Running multiple attention mechanisms in parallel, allowing the model to attend to different aspects of the input simultaneously.

An attention mechanism where each position in a sequence attends to all other positions, capturing relationships within the input.

Using knowledge learned from one task to improve performance on a related task. Foundation of modern LLM training.

The initial training phase where a model learns general language understanding from large amounts of unlabeled text.

The process of using a trained model to make predictions or generate outputs. Distinct from training.

The time delay between sending a request and receiving a response. Critical for real-time applications.

The number of requests or tokens a system can process per unit of time. Important for scaling.

Processing multiple requests together to improve throughput and GPU utilization.

Caching key-value pairs from attention layers to avoid recomputing them during autoregressive generation, speeding up inference.

An optimized attention algorithm that reduces memory usage and speeds up training/inference by reordering operations.

Training a smaller 'student' model to mimic a larger 'teacher' model, retaining most performance with fewer parameters.

A security vulnerability where malicious users craft prompts to override system instructions or extract sensitive information.

Techniques to bypass safety guardrails and make models generate prohibited content.

Connecting model outputs to factual sources or real-world data to reduce hallucinations.

The ability of LLMs to interact with external tools, APIs, or functions to extend their capabilities beyond text generation.

A prompting framework where models alternate between reasoning about what to do and taking actions with tools.