AAIA key terms

A training approach in which a model learns from labeled examples, pairing each input with the correct output so it can predict labels for new, unseen inputs.

A training approach that finds structure in unlabeled data, such as grouping similar records or reducing dimensionality, without any target answer being supplied.

The known target value or category attached to a training example, which the model is trained to reproduce or predict for new data.

An individual measurable input variable used by a model, such as a customer's age or a word count, often assembled with others into a feature vector.

The practice of selecting, transforming, and combining raw data into informative input variables so a model can learn more effectively from limited data.

A failure mode where a model memorizes training data, including noise, and performs well on seen examples but poorly on new ones.

A technique that penalizes model complexity, for example by shrinking weights, to reduce overfitting and improve generalization to unseen data.

The division of a dataset into portions used to fit a model and to estimate its performance on data it has not seen during training.

A setting chosen before training, such as learning rate or tree depth, that controls how the learning algorithm behaves rather than being learned from data.

An unsupervised technique that groups data points so members of the same group are more similar to each other than to points in other groups.

Two sources of prediction error: bias from overly simple assumptions causing systematic error, and variance from excessive sensitivity to training data causing instability.

A modeling mistake in which information unavailable at prediction time, such as target-derived or future data, contaminates training and inflates measured performance.

A model composed of layers of connected units that apply weighted sums and nonlinear activations, learning hierarchical representations from data through gradient-based optimization.

Machine learning using neural networks with many stacked layers, which learn increasingly abstract representations and typically require large datasets and substantial compute.

A training paradigm in which an agent learns a policy by taking actions in an environment and adjusting behavior based on rewards and penalties received over time.

A large model pretrained on broad data that can be adapted to many downstream tasks through prompting or fine-tuning rather than being built task by task.

A foundation model trained on vast text to predict and generate language, capable of tasks such as summarization, translation, question answering, and code generation.

A neural architecture built on self-attention that models relationships between all positions in a sequence in parallel, forming the basis of most modern language models.

A mechanism that computes weighted relevance between sequence positions, letting a model focus on the most informative parts of its input when producing each output.

A unit of text, often a word fragment, produced by a tokenizer; models process and generate sequences of tokens, and usage is frequently billed per token.

A dense numeric vector representing text, images, or other items so that semantically similar items lie close together in vector space, enabling search and comparison.

The maximum number of tokens a model can consider at once across its input and generated output, bounding how much material fits in a single request.

The instructions and input text supplied to a generative model that steer what output it produces, ranging from a simple question to a detailed multi-part template.

The practice of designing, structuring, and iterating on prompts, including instructions and examples, to reliably obtain accurate and well-formatted model outputs.

Including a small number of worked examples in a prompt so the model can infer the desired task, format, and style before responding to the real input.

A sampling setting that controls output randomness: lower values make responses more deterministic and focused, while higher values increase variety and creativity.

A confident model output that is fabricated or unsupported by evidence, such as invented citations or facts, arising because generation favors plausibility over verified truth.

Further training of a pretrained model on a smaller, task-specific dataset so it adapts to a domain, style, or narrow behavior beyond what prompting achieves.

A post-training method that uses human preference comparisons to train a reward model, which then guides the language model toward more helpful and safer responses.

A pattern that retrieves relevant documents from an external source and inserts them into the prompt so the model answers using current, verifiable context.

A data store optimized for similarity search over embeddings, returning items whose vectors are nearest to a query vector, commonly used to power retrieval.

Tying model output to authoritative sources or supplied context so claims can be traced and verified, reducing fabrication and improving trustworthiness in applications.

A system that uses a model to plan and take actions, such as calling tools or APIs, observing results, and iterating toward a goal with limited human involvement.

The capability of a model to invoke external functions, APIs, or services to fetch data or perform actions it cannot do through text generation alone.

A control that constrains model behavior or output, such as content filters, topic restrictions, or schema validation, to keep applications within intended and safe bounds.

Adversarial testing in which teams deliberately probe a system for harmful, biased, or unsafe behavior in order to find and fix weaknesses before deployment.

Arrangements ensuring people can monitor, review, intervene in, or override automated decisions, especially where errors could cause significant harm or loss of rights.

A short document describing a model's intended uses, training data, performance, and known limitations, helping downstream users judge whether it suits their context.

An attack in which malicious text, often hidden in retrieved content, manipulates a model into ignoring its instructions or leaking data or capabilities.

The documented origin and history of data, including its sources and transformations, used to assess legality, quality, and suitability for training or retrieval.