ACP-220 key terms

A neural network trained on large text corpora to predict and generate language, capable of tasks like summarization, translation, and question answering without task-specific retraining.

The basic unit of text an LLM reads or writes, roughly a word fragment. Billing, context limits, and latency are usually measured in tokens rather than characters.

The maximum number of tokens a model can consider at once, covering system instructions, conversation history, retrieved documents, and the generated response.

A sampling parameter that scales randomness in token selection. Lower values make output more deterministic and focused; higher values increase diversity and creative variation.

A decoding strategy that restricts choices to the smallest set of tokens whose cumulative probability exceeds a threshold, trading off coherence against variety.

A model output that is fluent but factually wrong or unsupported by any source, often produced when the model lacks grounding data or is asked beyond its knowledge.

Tying model responses to supplied evidence, such as retrieved documents or tool results, so answers can be traced to a verifiable source instead of parametric memory.

Further training of a pretrained model on a smaller, task-specific dataset so its weights adapt to a domain, style, or output format.

A family of adaptation methods that freeze most model weights and train a small number of added or selected parameters, cutting cost and storage versus full fine-tuning.

A PEFT technique that inserts small trainable low-rank matrices into model layers, letting a base model gain new behavior while the original weights stay unchanged.

When a model memorizes training examples too closely and performs poorly on unseen data, often from too little data, too many epochs, or excessive capacity.

Reducing the numeric precision of model weights and activations, such as from 16-bit to 8-bit or 4-bit, to shrink memory use and speed inference at some accuracy cost.

The process of running a trained model on new input to produce predictions or generated text, as opposed to the training phase that adjusts weights.

A dense numeric vector representing text, images, or other content so that semantic similarity can be measured by distance in vector space.

The practice of designing instructions, examples, and context formatting to steer a model toward accurate, well-structured, and safe outputs reliably.

A high-priority instruction set that defines a model application's role, tone, constraints, and safety rules before any user turns are processed.

Including a handful of worked input-output examples in the prompt so the model can infer the desired format and reasoning pattern without weight updates.

A prompting approach that asks the model to reason through intermediate steps before answering, improving performance on multi-step arithmetic, logic, and analysis tasks.

An architecture that retrieves relevant documents from a knowledge store and inserts them into the prompt so the model answers from fresh, domain-specific evidence.

A data store built to index embeddings and return nearest neighbors quickly, forming the retrieval backbone of semantic search and RAG pipelines.

Splitting source documents into smaller passages before embedding and indexing, balancing retrieval precision against loss of surrounding context.

Retrieval based on meaning rather than exact keyword matches, ranking results by vector similarity between the query and stored content.

A model-driven component that plans, chooses tools, and iterates on results to accomplish a goal, rather than producing a single one-shot response.

A mechanism letting a model request external functions or APIs with structured arguments, whose results are fed back so the model can continue reasoning.

Coordinating multiple model calls, retrieval steps, tools, and branching logic into a reliable application flow, including retries, state, and error handling.

A policy layer that inspects inputs and outputs to block harmful content, enforce topic limits, or prevent unsafe actions before a response reaches users.

An attack where untrusted content or user text tricks a model into ignoring its instructions, leaking data, or invoking tools improperly.

The labeled or curated examples used to adjust model weights, whose quality, coverage, and balance strongly determine downstream model behavior.

A held-out portion of data used to tune hyperparameters and detect overfitting during training without contaminating the final evaluation.

Systematically measuring a model's quality on defined tasks and metrics, such as accuracy, relevance, or safety, to guide release and comparison decisions.

A standardized task set and scoring method used to compare models or track regressions consistently across versions and configurations.

A training setting chosen before learning begins, such as learning rate, batch size, or epoch count, that shapes how weights are optimized.

Publishing a trained model to a serving environment where it accepts requests, with configuration for scaling, versioning, and access control.

A network address exposing a deployed model or service for requests, typically with authentication, rate limits, and autoscaling behind it.

A catalog tracking model versions, artifacts, metadata, and lineage so teams can promote, roll back, and audit models through their lifecycle.

Assigning identifiable versions to model artifacts and configurations so behavior changes can be traced, compared, and reverted safely.

A tool embedded in a developer environment that suggests code, completes functions, explains snippets, or generates tests from natural-language prompts.

Predicting the next lines or tokens of code from surrounding context, used to accelerate routine editing while requiring developer review.

A model that accepts or produces more than one data type, such as text alongside images, audio, or video, enabling cross-modal understanding and generation.

A multimodal model jointly trained on images and text that can describe pictures, answer questions about visuals, or follow instructions involving visual content.

Converting text in scanned images or photos into machine-readable characters, often paired with layout analysis to preserve reading order.

Automatic speech recognition that transcribes spoken audio into written text, handling variation in accent, noise, and speaking rate.

Synthesizing natural-sounding spoken audio from written text, with controls for voice, speed, and pronunciation.