\[Live session\] Choosing safer LLMs: From LLM benchmarks to your production agents 🚀

July 21, 2026 \| 5PM CEST

1

Days

:

00

Hours

:

26

Minutes

:

16

Seconds

[Save your spot](https://watch.getcontrast.io/register/giskard-live-session-choosing-safer-llms-from-llm-benchmarks-to-your-production-agents?utm_source=website&utm_medium=banner)

🎥 Discover how to red team your AI agents.

[Watch demo](/content/giskard-hub-demo/index.html)

.svg)\\
\\
\\
5,660](https://github.com/Giskard-AI/giskard)

# Glossary

Dive into essential terms curated by AI quality, security & compliance experts. Gain clarity in the new language of AI.

Thank you! Your submission has been received!

Oops! Something went wrong while submitting the form.

Filter alphabetical

A

B

C

D

E

F

G

H

I

J

K

L

M

N

O

P

Q

R

S

T

U

V

W

X

Y

Z

Letter

## Abductive Logic Programming

ALP uses abduction to hypothesize explanations for observations when knowledge is incomplete.

[Learn more](/content/glossary/abductive-logic-programming/index.html)

A

## Abstract Data Type

ADTs specify what operations a type supports, not how those operations are implemented.

[Learn more](/content/glossary/abstract-data-type/index.html)

A

## Accuracy Metric

Accuracy measures the proportion of correct predictions; use with care on imbalanced data.

[Learn more](/content/glossary/accuracy-metric/index.html)

A

## ACID Transactions

ACID properties ensure database transactions are reliable and consistent under concurrent load.

[Learn more](/content/glossary/acid-transactions/index.html)

A

## Activation Functions

Activation functions decide whether neurons fire and enable nonlinear learning in neural nets.

[Learn more](/content/glossary/activation-functions/index.html)

A

## Active Learning in Machine Learning

Active learning selects which data to label next so models improve with fewer annotations.

[Learn more](/content/glossary/active-learning-in-machine-learning/index.html)

A

## Adaptive Gradient Algorithm (AdaGrad)

AdaGrad adjusts learning rates per parameter based on accumulated gradient history.

[Learn more](/content/glossary/adaptive-gradient-algorithm-adagrad/index.html)

A

## Adversarial Machine Learning

Adversarial machine learning covers attacks that fool models and defenses that improve robustness against evasion and poisoning.

[Learn more](/content/glossary/adversarial-machine-learning/index.html)

H

## Agent2Agent Protocol

A2A is an open protocol for agents to discover peers, negotiate tasks, and exchange results securely.

[Learn more](/content/glossary/agent2agent-protocol/index.html)

A

## AgentBench Agent Benchmark

AgentBench measures how well language-model agents handle multi-step tasks across diverse environments.

[Learn more](/content/glossary/agentbench-agent-benchmark/index.html)

A

## AgentHarm Safety Benchmark

AgentHarm tests AI agents on multi-step tasks that probe safety boundaries and harm prevention.

[Learn more](/content/glossary/agentharm-safety-benchmark/index.html)

A

## Agentic Chunking

Agentic chunking lets a language model segment content by semantics for smarter retrieval.

[Learn more](/content/glossary/agentic-chunking/index.html)

A

## Agentic Orchestration

Agentic orchestration coordinates decisions and actions among humans, agents, and automation layers.

[Learn more](/content/glossary/agentic-orchestration/index.html)

A

## Agentic RAG

Agentic RAG lets agents actively plan retrieval and tool use beyond a single static RAG pass.

[Learn more](/content/glossary/agentic-rag/index.html)

A

## Agentic Workflow

Agentic workflows use autonomous agents that iterate through planning, tool use, and observation.

[Learn more](/content/glossary/agentic-workflow/index.html)

A

## Agent Observability

Agent observability systematically records runtime signals—prompts, plans, tools, and side effects—for reliable AI agents.

[Learn more](/content/glossary/agent-observability/index.html)

## AI Agent

AI agents autonomously perceive, decide, and act to achieve goals in an environment.

[Learn more](/content/glossary/ai-agent/index.html)

A

## AI Agent Evaluation

AI agent evaluation is the generate→score→compare→log→improve loop for trustworthy agent releases.

[Learn more](/content/glossary/ai-agent-evaluation/index.html)

A

## AI Agent Observability

AI agent observability turns opaque agent decisions into measurable, auditable production signals.

[Learn more](/content/glossary/ai-agent-observability/index.html)

A

## AI Center of Excellence (AI CoE)

An AI CoE centralizes expertise to guide AI development, governance, and rollout enterprise-wide.

[Learn more](/content/glossary/ai-center-of-excellence-ai-coe/index.html)

A

## AI Content Moderation

AI content moderation automatically screens user-generated content for policy violations and safety risks.

[Learn more](/content/glossary/ai-content-moderation/index.html)

A

## AI Copilots

AI copilots are virtual assistants that boost productivity by automating and guiding tasks in context.

[Learn more](/content/glossary/ai-copilots/index.html)

A

## AI Data Labeling

Data labeling marks inputs and outputs so supervised ML models can learn from examples.

[Learn more](/content/glossary/ai-data-labeling/index.html)

A

## AI Fairness

AI fairness seeks to prevent biased outcomes and promote equitable treatment in automated decisions.

[Learn more](/content/glossary/ai-fairness/index.html)

A

## AI Model Validation

Model validation confirms an AI system is accurate, reliable, and secure enough for its intended use.

[Learn more](/content/glossary/ai-model-validation/index.html)

A

## AI Observability

AI observability provides ongoing visibility into model performance and behavior in production.

[Learn more](/content/glossary/ai-observability/index.html)

A

## AI Steerability

Steerability means fine-grained control over how AI models behave relative to policies and goals.

[Learn more](/content/glossary/ai-steerability/index.html)

A

## Alignment Metric (NLI)

NLI-based alignment scores whether a text entails, contradicts, or is neutral to a reference.

[Learn more](/content/glossary/alignment-metric-nli/index.html)

## AlpacaEval

AlpacaEval automates evaluation of instruction-following LLMs with toolkits and public leaderboards.

[Learn more](/content/glossary/alpacaeval/index.html)

A

## AlpacaEval Conversation Benchmark

AlpacaEval’s conversation benchmark tests instruction following and response appropriateness.

[Learn more](/content/glossary/alpacaeval-conversation-benchmark/index.html)

A

## ANFIS

ANFIS merges neural nets and fuzzy inference to model complex nonlinear relationships from data.

[Learn more](/content/glossary/anfis/index.html)

A

## Anomaly Detection

Anomaly detection identifies unusual patterns that may signal errors, fraud, or attacks.

[Learn more](/content/glossary/anomaly-detection/index.html)

A

## Answer Relevancy Metric

Answer Relevancy scores how well a generated answer addresses the input prompt.

[Learn more](/content/glossary/answer-relevancy-metric/index.html)

## Answer Relevancy RAG Metric

In RAG systems, this metric scores how well answers align with the asked question.

[Learn more](/content/glossary/answer-relevancy-rag-metric/index.html)

## APPS Coding Benchmark

APPS evaluates how well LLMs solve programming problems with correct, efficient code.

[Learn more](/content/glossary/apps-coding-benchmark/index.html)

A

## ARC Reasoning Benchmark

ARC tests whether language models reason through grade-school science questions.

[Learn more](/content/glossary/arc-reasoning-benchmark/index.html)

A

## Argument Correctness Metric

This metric scores logical validity and structure of arguments produced by AI agents.

[Learn more](/content/glossary/argument-correctness-metric/index.html)

## Artificial Neural Network

ANNs learn from data via interconnected units that approximate complex input–output relationships.

[Learn more](/content/glossary/artificial-neural-network/index.html)

A

## ASCII Smuggling Injection Attack

ASCII smuggling injects hidden instructions via invisible characters that models still parse.

[Learn more](/content/glossary/ascii-smuggling-injection-attack/index.html)

## Attention in Machine Learning

Attention lets models focus on the most relevant parts of an input when making predictions.

[Learn more](/content/glossary/attention-in-machine-learning/index.html)

A

## Attribute

Attributes are the data fields or features that models use to learn patterns and make predictions.

[Learn more](/content/glossary/attribute/index.html)

A

## Auto-Encoders

Autoencoders learn compact representations by encoding data and decoding it back to the original.

[Learn more](/content/glossary/auto-encoders/index.html)

A

## Automated Machine Learning

AutoML streamlines model development by automating search over pipelines and hyperparameters.

[Learn more](/content/glossary/automated-machine-learning/index.html)

A

## AutoML

AutoML reduces manual ML work by automating model search, tuning, and basic evaluation.

[Learn more](/content/glossary/automl/index.html)

A

## Autonomous Agents

Autonomous agents perceive context, set goals, and act with little ongoing human control.

[Learn more](/content/glossary/autonomous-agents/index.html)

A

## Autoregressive Model

Autoregressive models forecast future points from lagged past values in a sequence.

[Learn more](/content/glossary/autoregressive-model/index.html)

A

## Average Precision

Average Precision averages precision across recall levels, often at multiple IoU thresholds for detection.

[Learn more](/content/glossary/average-precision/index.html)

A

## AWS Bedrock

AWS Bedrock provides API access to multiple foundation models with enterprise security features.

[Learn more](/content/glossary/aws-bedrock/index.html)

A

## AWS Sagemaker

SageMaker provides IDE, training, and deployment tools for cloud machine learning on AWS.

[Learn more](/content/glossary/aws-sagemaker/index.html)

A

## Backpropagation

Backpropagation calculates gradients through a network so weights can be updated during learning.

[Learn more](/content/glossary/backpropagation/index.html)

B

## Backpropagation Algorithm

Backpropagation adjusts weights in reverse through the network to minimize prediction error.

[Learn more](/content/glossary/backpropagation-algorithm/index.html)

B

## Bagging in Machine Learning

Bagging builds an ensemble from bootstrap samples to stabilize predictions and cut variance.

[Learn more](/content/glossary/bagging-in-machine-learning/index.html)

B

## Baseline Distribution

Baseline distributions provide the minimum performance bar for evaluating advanced models.

[Learn more](/content/glossary/baseline-distribution/index.html)

B

## Baseline Models

Baseline models are simple starting points used to judge whether complex models add real value.

[Learn more](/content/glossary/baseline-models/index.html)

B

## Batch Normalization

Batch normalization stabilizes training by normalizing interlayer activations within mini-batches.

[Learn more](/content/glossary/batch-normalization/index.html)

B

## Batch Standardization

Batch standardization normalizes outputs between layers to mitigate internal covariate shift.

[Learn more](/content/glossary/batch-standardization/index.html)

B

## Bayes' theorem

Bayes' theorem computes conditional probability from priors and likelihoods.

[Learn more](/content/glossary/bayes-theorem/index.html)

B

## Berkeley Function-Calling Leaderboard Domain-Specific Benchmark

Domain benchmark for multi-language LLM function calling, parallel calls, and relevance detection.

[Learn more](/content/glossary/berkeley-function-calling-leaderboard-domain-specific-benchmark/index.html)

B

## BERT

Bidirectional transformer encoder for contextual NLP representations and fine-tuning.

[Learn more](/content/glossary/bert/index.html)

B

## Best-of-N Prompt Injection Attack

Adversarial method that samples many prompt variants and keeps any that bypass injection defenses.

[Learn more](/content/glossary/best-of-n-prompt-injection-attack/index.html)

## Bias Metric

Metric for political, gender, and social bias in model outputs used in decisions.

[Learn more](/content/glossary/bias-metric/index.html)

## Bias Variance Tradeoff

ML tradeoff between underfitting bias and overfitting variance.

[Learn more](/content/glossary/bias-variance-tradeoff/index.html)

B

## BigBench Reasoning Benchmark

Collaborative benchmark covering logic, math, and language comprehension for LLMs.

[Learn more](/content/glossary/bigbench-reasoning-benchmark/index.html)

B

## Binary Classification

Supervised task assigning each input to one of two classes.

[Learn more](/content/glossary/binary-classification/index.html)

B

## Binary Cross Entropy

Log-loss metric for binary probability predictions versus true labels.

[Learn more](/content/glossary/binary-cross-entropy/index.html)

B

## Binomial Distribution

Probability distribution of successes in n identical independent Bernoulli trials.

[Learn more](/content/glossary/binomial-distribution/index.html)

B

## Black Box Model

Models whose internal decision mechanics are opaque to users and auditors.

[Learn more](/content/glossary/black-box-model/index.html)

B

## BLEU

Automatic metric comparing generated text to references via n-gram precision.

[Learn more](/content/glossary/bleu/index.html)

B

## Blue Teaming Capabilities (OWASP Taxonomy)

Defense capabilities: guardrails, runtime firewalls, AI-SPM, and detection that blocks or catches attacks.

[Learn more](/content/glossary/blue-teaming-capabilities-owasp-taxonomy/index.html)

B

## Broken Function Level Authorization Excessive Agency Attack

Probe that tests whether agents execute functions beyond their authorized privilege level.

[Learn more](/content/glossary/broken-function-level-authorization-excessive-agency-attack/index.html)

## Broken Object Level Authorization Excessive Agency Attack

Probe that checks whether agents fetch unauthorized objects via prompts or tool calls.

[Learn more](/content/glossary/broken-object-level-authorization-excessive-agency-attack/index.html)

## Calibration Curve

Plot comparing predicted class probabilities to observed frequencies.

[Learn more](/content/glossary/calibration-curve/index.html)

C

## Canonical Schema

Standardized data model enabling consistent exchange across multiple systems.

[Learn more](/content/glossary/canonical-schema/index.html)

C

## Catastrophic Forgetting

Neural-net phenomenon where new learning overwrites previously acquired capabilities.

[Learn more](/content/glossary/catastrophic-forgetting/index.html)

C

## Catboost

Gradient boosting toolkit specializing in categorical feature handling.

[Learn more](/content/glossary/catboost/index.html)

C

## Categorical Variables

Features that take values from a finite set of named categories.

[Learn more](/content/glossary/categorical-variables/index.html)

C

## Causal Language Modeling (CLM)

Autoregressive training that predicts the next token from left context only.

[Learn more](/content/glossary/causal-language-modeling-clm/index.html)

C

## CBRN Harmful Content Attack

Adversarial probes seeking CBRN weapons-related assistance from AI models.

[Learn more](/content/glossary/cbrn-harmful-content-attack/index.html)

## Chain-of-Thought

Intermediate step-by-step reasoning used by LLMs; also a security attack surface.

[Learn more](/content/glossary/chain-of-thought/index.html)

C

## Chain-of-Thought Evaluation Metric

Metric that scores model outputs using systematic step-by-step reasoning criteria.

[Learn more](/content/glossary/chain-of-thought-evaluation-metric/index.html)

## Chain-of-Thought Prompting

Prompting method that elicits intermediate reasoning steps for complex tasks.

[Learn more](/content/glossary/chain-of-thought-prompting/index.html)

C

## Chatbot Arena Conversation Benchmark

Human preference benchmark ranking chat models with pairwise votes and Elo scores.

[Learn more](/content/glossary/chatbot-arena-conversation-benchmark/index.html)

C

## ChatGLM

Family of bilingual conversational LLMs focused on Chinese–English dialogue.

[Learn more](/content/glossary/chatglm/index.html)

C

## CI CD for Machine Learning

Applying continuous integration and delivery to ML training and deployment.

[Learn more](/content/glossary/ci-cd-for-machine-learning/index.html)

C

## Citation Framing Injection Attack

Prompt injection that frames malicious asks as academic citations or scholarly references.

[Learn more](/content/glossary/citation-framing-injection-attack/index.html)

## Classification Threshold

Cutoff on predicted probability used to assign a hard class label.

[Learn more](/content/glossary/classification-threshold/index.html)

C

## Class Imbalance

Training data where one class far outnumbers others, skewing naive metrics.

[Learn more](/content/glossary/class-imbalance/index.html)

C

## Clustering Algorithms

Unsupervised methods that group data points by similarity without labels.

[Learn more](/content/glossary/clustering-algorithms/index.html)

C

## CodeContests Coding Benchmark

Competitive-programming benchmark for assessing algorithmic code generation by LLMs.

[Learn more](/content/glossary/codecontests-coding-benchmark/index.html)

C

## Code Execution Metric

Metric that evaluates generated code by executing it against tests or expected outputs.

[Learn more](/content/glossary/code-execution-metric/index.html)

## Code Interpreter

LLM tool that executes model-generated code, typically inside a sandbox.

[Learn more](/content/glossary/code-interpreter/index.html)

C

## CodeXGLUE Coding Benchmark

Microsoft multi-task benchmark suite for code understanding and generation models.

[Learn more](/content/glossary/codexglue-coding-benchmark/index.html)

C

## CommonsenseQA Reasoning Benchmark

Multiple-choice QA benchmark requiring everyday commonsense reasoning.

[Learn more](/content/glossary/commonsenseqa-reasoning-benchmark/index.html)

C

## Competitors Brand Damage Attack

Probe that checks if AI can be pushed to endorse competitors or damage brand reputation.

[Learn more](/content/glossary/competitors-brand-damage-attack/index.html)

## Complex Event Processing

Real-time analysis of event streams to detect complex patterns and situations.

[Learn more](/content/glossary/complex-event-processing/index.html)

C

## Computer Vision

AI field for interpreting visual data such as images and video.

[Learn more](/content/glossary/computer-vision/index.html)

C

## Confusion Matrix in Machine Learning

Table summarizing classifier correct and incorrect predictions by class.

[Learn more](/content/glossary/confusion-matrix-in-machine-learning/index.html)

C

## Context Compliance Harmful Content Attack

Multi-turn probe using fabricated conversation history to elicit harmful content.

[Learn more](/content/glossary/context-compliance-harmful-content-attack/index.html)

## Contextual Precision Metric

RAG metric for whether relevant context ranks above irrelevant chunks.

[Learn more](/content/glossary/contextual-precision-metric/index.html)

## Contextual Recall Metric

RAG metric for whether necessary supporting information was retrieved.

[Learn more](/content/glossary/contextual-recall-metric/index.html)

## Contextual Relevancy Metric

Metric scoring how relevant retrieved or supplied context is to the query.

[Learn more](/content/glossary/contextual-relevancy-metric/index.html)

## Contextual Relevancy RAG Metric

RAG metric for on-topic retrieved documents before answer generation.

[Learn more](/content/glossary/contextual-relevancy-rag-metric/index.html)

## Context Window

Maximum token span an LLM can attend to in a single request.

[Learn more](/content/glossary/context-window/index.html)

C

## Continuous Integration Model

CI practice applied to model training, testing, and validation pipelines.

[Learn more](/content/glossary/continuous-integration-model/index.html)

C

## Continuous Validation

Repeated post-deploy testing of models for quality, safety, and drift.

[Learn more](/content/glossary/continuous-validation/index.html)

C

## Contrastive Learning

Self-supervised method learning by comparing similar vs dissimilar sample pairs.

[Learn more](/content/glossary/contrastive-learning/index.html)

C

## Conversational Agent

AI systems that hold multi-turn natural-language conversations with users.

[Learn more](/content/glossary/conversational-agent/index.html)

C

## Convex Optimization

Mathematical optimization where convexity yields tractable global solutions.

[Learn more](/content/glossary/convex-optimization/index.html)

C

## Convolutional Neural Networks

Neural nets with convolutional layers for spatial pattern recognition.

[Learn more](/content/glossary/convolutional-neural-networks/index.html)

C

## Copyright Violations Harmful Content Attack

Probe for unauthorized reproduction or distribution of copyrighted materials by AI.

[Learn more](/content/glossary/copyright-violations-harmful-content-attack/index.html)

## Corrective RAG

RAG pattern that grades context and triggers corrective retrieval when needed.

[Learn more](/content/glossary/corrective-rag/index.html)

C

## Correctness Metric

Metric comparing model outputs to expected ground-truth answers.

[Learn more](/content/glossary/correctness-metric/index.html)

## Crescendo Harmful Content Attack

Incremental multi-turn strategy that steers models toward harmful content gradually.

[Learn more](/content/glossary/crescendo-harmful-content-attack/index.html)

## Cross-Lingual Language Models

AI models that understand and generate text across multiple languages via shared representations.

[Learn more](/content/glossary/cross-lingual-language-models/index.html)

C

## Cross Session Leak Data Privacy Attack

Privacy probe testing whether sensitive data leaks across separate AI sessions.

[Learn more](/content/glossary/cross-session-leak-data-privacy-attack/index.html)

## Cross Validation Modeling

A model assessment method that tests on multiple data folds for more reliable generalization estimates.

[Learn more](/content/glossary/cross-validation-modeling/index.html)

C

## Custom LLM Judge Metric

An LLM-as-judge evaluation method with custom criteria and prompts for subjective quality scoring.

[Learn more](/content/glossary/custom-llm-judge-metric/index.html)

C

## CyberSecEval Harmful Content Attack

A red-team probe using CyberSecEval samples to check if models produce harmful cybersecurity content.

[Learn more](/content/glossary/cyberseceval-harmful-content-attack/index.html)

## DAN Prompt Injection Attack

DAN-style prompt injection uses role-play to push models past safety protocols as an unrestricted assistant.

[Learn more](/content/glossary/dan-prompt-injection-attack/index.html)

## Data Annotation in AI

Labeling data samples so supervised machine learning models can learn from them.

[Learn more](/content/glossary/data-annotation-in-ai/index.html)

D

## Data Augmentation

Techniques that synthetically expand training sets with modified copies of existing samples.

[Learn more](/content/glossary/data-augmentation/index.html)

D

## Data Binning

A preprocessing method that maps continuous numbers into discrete bin intervals.

[Learn more](/content/glossary/data-binning/index.html)

D

## Data-Centric AI

An approach that improves AI by systematically raising data quality and coverage.

[Learn more](/content/glossary/data-centric-ai/index.html)

D

## Data Cleaning

The process of fixing or removing inaccurate, duplicate, or malformed data before analysis.

[Learn more](/content/glossary/data-cleaning/index.html)

D

## Data Decomposition

Separating a time series into trend, seasonal, and residual components.

[Learn more](/content/glossary/data-decomposition/index.html)

D

## Data Flywheel

A loop where usage generates data that improves models and drives further usage.

[Learn more](/content/glossary/data-flywheel/index.html)

D

## Data Granularity

How fine-grained or aggregated the records in a dataset are.

[Learn more](/content/glossary/data-granularity/index.html)

D

## Data Logging

Automated collection and storage of time-stamped measurements or events.

[Learn more](/content/glossary/data-logging/index.html)

D

## Data Mart

A subject-specific subset of a data warehouse for a department or business function.

[Learn more](/content/glossary/data-mart/index.html)

D

## Data Purification

Cleaning inaccurate, corrupt, duplicate, or missing data from datasets.

[Learn more](/content/glossary/data-purification/index.html)

D

## Data Science Platform

Integrated software for collaborative data prep, modeling, and insight sharing.

[Learn more](/content/glossary/data-science-platform/index.html)

D

## Data Science Techniques

Analytical and modeling methods used to extract actionable insights from data.

[Learn more](/content/glossary/data-science-techniques/index.html)

D

## Data Science Tools

Software libraries and platforms used to collect, analyze, and model data.

[Learn more](/content/glossary/data-science-tools/index.html)

D

## Datasets And Machine Learning

Structured collections of examples that train, validate, and test ML models.

[Learn more](/content/glossary/datasets-and-machine-learning/index.html)

D

## Data Vault

A data warehouse modeling approach built around hubs, links, and satellites.

[Learn more](/content/glossary/data-vault/index.html)

D

## Data Versioning

Tracking and labeling dataset states so ML experiments remain reproducible.

[Learn more](/content/glossary/data-versioning/index.html)

D

## Data Visualizations

Graphical representation of data to reveal trends, anomalies, and patterns.

[Learn more](/content/glossary/data-visualizations/index.html)

D

## Decision Boundary

The region in feature space where a model's predicted class switches.

[Learn more](/content/glossary/decision-boundary/index.html)

D

## Decision Intelligence

A discipline that blends analytics, AI, and decision processes for better outcomes.

[Learn more](/content/glossary/decision-intelligence/index.html)

D

## Decision Tree

A model that routes examples through feature-based splits to a leaf prediction.

[Learn more](/content/glossary/decision-tree/index.html)

D

## Decision Tree In Machine Learning

Supervised models that learn hierarchical splits for classification or regression.

[Learn more](/content/glossary/decision-tree-in-machine-learning/index.html)

D

## Decomposed Evaluation Metric

An evaluation approach that scores separate quality dimensions instead of one opaque overall grade.

[Learn more](/content/glossary/decomposed-evaluation-metric/index.html)

D

## Deep Belief Networks

Generative deep models built by stacking restricted Boltzmann machines.

[Learn more](/content/glossary/deep-belief-networks/index.html)

D

## Deep Learning

Neural network methods with many layers that learn hierarchical representations.

[Learn more](/content/glossary/deep-learning/index.html)

D

## Deep learning Algorithms

Neural architectures and training methods used to build deep learning systems.

[Learn more](/content/glossary/deep-learning-algorithms/index.html)

D

## Deep Q-Network

A reinforcement learning method that uses deep nets to approximate action values.

[Learn more](/content/glossary/deep-q-network/index.html)

D

## Deep Reinforcement Learning

Using deep neural networks as function approximators inside reinforcement learning.

[Learn more](/content/glossary/deep-reinforcement-learning/index.html)

D

## Deepset Injection Attack

A prompt-injection evaluation that probes whether models follow attacker instructions over system intent.

[Learn more](/content/glossary/deepset-injection-attack/index.html)

## Deep SHAP

A SHAP-based method for explaining predictions of deep neural networks.

[Learn more](/content/glossary/deep-shap/index.html)

D

## Degradation Model

Statistical or ML models of how performance decays under time or stress.

[Learn more](/content/glossary/degradation-model/index.html)

D

## Denial of Service Attack

DoS against AI systems via long outputs, loops, or resource-heavy prompts that exhaust capacity or spend.

[Learn more](/content/glossary/denial-of-service-attack/index.html)

## DenseNET

A convolutional architecture with dense skip connections between layers.

[Learn more](/content/glossary/densenet/index.html)

D

## Density Based Clustering

Clustering that finds dense regions of arbitrary shape and marks outliers as noise.

[Learn more](/content/glossary/density-based-clustering/index.html)

D

## Diffusion Models

Generative models that create samples by reversing a stepwise noising process.

[Learn more](/content/glossary/diffusion-models/index.html)

D

## Dimensionality Reduction

Techniques that map high-dimensional data into a lower-dimensional representation.

[Learn more](/content/glossary/dimensionality-reduction/index.html)

D

## Direct Preference Optimization

An alignment method that trains LLMs from human preference pairs without a separate reward model.

[Learn more](/content/glossary/direct-preference-optimization/index.html)

D

## Divergent Repetitions Training Data Extraction Attack

An extraction probe that uses divergent repetition patterns to surface memorized or sensitive training content.

[Learn more](/content/glossary/divergent-repetitions-training-data-extraction-attack/index.html)

## DoNotAnswer Harmful Content Attack

A harmful-content probe based on DoNotAnswer-style questions that safe AI systems should refuse.

[Learn more](/content/glossary/donotanswer-harmful-content-attack/index.html)

## Dplyr

An R package for expressive, pipe-friendly tabular data manipulation.

[Learn more](/content/glossary/dplyr/index.html)

D

## Drift Monitoring

Tracking distribution or behavior shifts in data and models against a known baseline.

[Learn more](/content/glossary/drift-monitoring/index.html)

D

## DROP Reasoning Benchmark

A QA benchmark testing discrete reasoning skills over paragraph-length contexts.

[Learn more](/content/glossary/drop-reasoning-benchmark/index.html)

D

## Early Stopping

A regularization tactic that stops training when held-out metrics plateau or worsen.

[Learn more](/content/glossary/early-stopping/index.html)

E

## Embedding Projector

Interactive tools for projecting and exploring embedding spaces visually.

[Learn more](/content/glossary/embedding-projector/index.html)

E

## Embedding Similarity Metric

A metric that scores how close two texts are in embedding space for semantic evaluation.

[Learn more](/content/glossary/embedding-similarity-metric/index.html)

## Embeddings in Machine Learning

Dense vector representations that encode semantic or structural similarity for ML tasks.

[Learn more](/content/glossary/embeddings-in-machine-learning/index.html)

E

## Embodied Agents

AI agents that perceive and act through a physical or simulated body, not text alone.

[Learn more](/content/glossary/embodied-agents/index.html)

E

## Emotion Detection Metric

A metric that classifies or scores emotional tone in model outputs for UX and safety reviews.

[Learn more](/content/glossary/emotion-detection-metric/index.html)

## Encoding Prompt Injection Attack

Prompt injection that obfuscates harmful instructions via encodings like Base64 or NATO phonetic text.

[Learn more](/content/glossary/encoding-prompt-injection-attack/index.html)

## End-to-End Evaluation

Testing the complete AI workflow from user input through tools and final response quality.

[Learn more](/content/glossary/end-to-end-evaluation/index.html)

E

## Ensemble Learning

Methods that aggregate multiple models to produce a stronger predictor.

[Learn more](/content/glossary/ensemble-learning/index.html)

E

## Enterprise Generative AI

Organization-scale generative AI with governance, security controls, and continuous evaluation.

[Learn more](/content/glossary/enterprise-generative-ai/index.html)

E

## Epoch in machine Learning

One full pass of the training set through the learning algorithm.

[Learn more](/content/glossary/epoch-in-machine-learning/index.html)

E

## Ethical AI

Practices that align AI systems with fairness, transparency, privacy, and human rights.

[Learn more](/content/glossary/ethical-ai/index.html)

E

## ETL Pipeline

Extract–transform–load workflow moving data from sources into analytics systems.

[Learn more](/content/glossary/etl-pipeline/index.html)

E

## Evolutionary Algorithms

Population-based optimizers that evolve candidate solutions via selection and variation.

[Learn more](/content/glossary/evolutionary-algorithms/index.html)

E

## Exact Match Metric

A precise evaluation method that scores answers via direct string equality to gold labels.

[Learn more](/content/glossary/exact-match-metric/index.html)

## Explainable AI (XAI)

AI designed to provide understandable insights into how and why decisions are made.

[Learn more](/content/glossary/explainable-ai-xai/index.html)

E

## Exploratory Data Analysis (EDA)

Foundational analysis that reveals dataset structure, often with visualizations.

[Learn more](/content/glossary/exploratory-data-analysis-eda/index.html)

E

## F1-Score Metric

Harmonic mean of precision and recall for evaluating classification and extraction quality.

[Learn more](/content/glossary/f1-score-metric/index.html)

## Facial Recognition

Biometric tech that matches faces in images or video to known identities.

[Learn more](/content/glossary/facial-recognition/index.html)

F

## Failure Analysis Machine Learning

Systematic diagnosis of why machine learning models make incorrect predictions.

[Learn more](/content/glossary/failure-analysis-machine-learning/index.html)

F

## Faithfulness Metric

Scores how well an answer stays faithful to retrieved or provided source context.

[Learn more](/content/glossary/faithfulness-metric/index.html)

## False Positive Rate

Proportion of actual negatives incorrectly classified as positive.

[Learn more](/content/glossary/false-positive-rate/index.html)

F

## Feature Engineering

Creating and transforming input variables to improve machine learning models.

[Learn more](/content/glossary/feature-engineering/index.html)

F

## Feature Selection

Choosing the most informative subset of features for a model.

[Learn more](/content/glossary/feature-selection/index.html)

F

## Feature Vector

An ordered set of numeric features representing a single data example.

[Learn more](/content/glossary/feature-vector/index.html)

F

## Federated Learning

Collaborative training that keeps raw data local and aggregates model updates.

[Learn more](/content/glossary/federated-learning/index.html)

F

## Feedback Loop

When model predictions alter future data and can amplify bias or error.

[Learn more](/content/glossary/feedback-loop/index.html)

F

## Few-Shot Evaluation Metric

Assesses model performance when learning from a small number of in-context examples.

[Learn more](/content/glossary/few-shot-evaluation-metric/index.html)

## Few-Shot Learning

ML paradigm where models adapt to new tasks from only a few labeled examples.

[Learn more](/content/glossary/few-shot-learning/index.html)

F

## Few-Shot Prompting

Prompting technique that includes a few demonstrations to guide LLM behavior.

[Learn more](/content/glossary/few-shot-prompting/index.html)

F

## FinBen Domain-Specific Benchmark

A domain benchmark for assessing LLM performance on financial tasks.

[Learn more](/content/glossary/finben-domain-specific-benchmark/index.html)

F

## Folium

Python wrapper for creating interactive Leaflet maps from geospatial data.

[Learn more](/content/glossary/folium/index.html)

F

## F-score

A performance metric combining precision and recall into a single score (often F1).

[Learn more](/content/glossary/f-score/index.html)

F

## GAIA Agent Benchmark

A benchmark of real-world questions for evaluating general AI assistants and agents.

[Learn more](/content/glossary/gaia-agent-benchmark/index.html)

G

## Gaussian Distribution

The normal bell-curve distribution parameterized by mean and variance.

[Learn more](/content/glossary/gaussian-distribution/index.html)

G

## Gaussian Mixture Model

Probabilistic model representing data as a mixture of Gaussian components.

[Learn more](/content/glossary/gaussian-mixture-model/index.html)

G

## GCG Injection Harmful Content Attack

Greedy Coordinate Gradient (GCG) attacks optimize adversarial suffixes that jailbreak LLM safety filters.

[Learn more](/content/glossary/gcg-injection-harmful-content-attack/index.html)

## Generalist Language Model

An LLM designed for broad tasks across domains rather than one specialized use case.

[Learn more](/content/glossary/generalist-language-model/index.html)

G

## Generalized Linear Models

A statistical family linking linear predictors to diverse response distributions.

[Learn more](/content/glossary/generalized-linear-models/index.html)

G

## Generative Adversarial Networks

Two-network architecture where a generator learns by fooling a discriminator.

[Learn more](/content/glossary/generative-adversarial-networks/index.html)

G

## Generative Agents

LLM-based agents that simulate believable behavior via memory, planning, and reflection.

[Learn more](/content/glossary/generative-agents/index.html)

G

## G-Eval

An LLM-as-judge method that scores outputs via structured evaluation rubrics.

[Learn more](/content/glossary/g-eval/index.html)

G

## GOAT Attack Harmful Content Attack

GOAT automates multi-turn adversarial prompting to uncover jailbreaks and harmful-content failures in LLMs.

[Learn more](/content/glossary/goat-attack-harmful-content-attack/index.html)

## Gradient Boosting

Sequential ensemble method that adds weak learners to correct prior residuals.

[Learn more](/content/glossary/gradient-boosting/index.html)

G

## Gradient Descent in Machine Learning

Iterative optimization that follows gradients to minimize training loss.

[Learn more](/content/glossary/gradient-descent-in-machine-learning/index.html)

G

## Grandma Framing Injection Attack

Emotional framing that wraps harmful asks as innocent family scenarios to jailbreak LLMs.

[Learn more](/content/glossary/grandma-framing-injection-attack/index.html)

## Grid Search

Brute-force search over a specified hyperparameter grid.

[Learn more](/content/glossary/grid-search/index.html)

G

## Groundedness Metric

Evaluates whether answer claims are supported by provided or retrieved context.

[Learn more](/content/glossary/groundedness-metric/index.html)

## Ground Truth

Authoritative reference data against which model predictions are scored.

[Learn more](/content/glossary/ground-truth/index.html)

G

## Grouped Query Attention

Transformer attention variant that groups queries to reduce KV cache size.

[Learn more](/content/glossary/grouped-query-attention/index.html)

G

## GSM8K Math Benchmark

Benchmark of grade-school math word problems for multi-step reasoning.

[Learn more](/content/glossary/gsm8k-math-benchmark/index.html)

G

## Hallucination Index

A composite score for how often an LLM fabricates facts without support.

[Learn more](/content/glossary/hallucination-index/index.html)

H

## Hallucination Metric

An evaluation measure for unsupported or fabricated claims in LLM outputs.

[Learn more](/content/glossary/hallucination-metric/index.html)

## Handling Outliers

Methods to detect and treat extreme values that distort analysis or training.

[Learn more](/content/glossary/handling-outliers/index.html)

H

## HarmBench Harmful Content Attack

A benchmark suite for measuring LLM resistance to harmful content generation.

[Learn more](/content/glossary/harmbench-harmful-content-attack/index.html)

## Hash tables

Fast key–value maps implemented with hashing and collision resolution.

[Learn more](/content/glossary/hash-tables-9fe0a/index.html)

H

## Hash Tables

Data structure mapping keys to values via hashing for fast average-case lookup.

[Learn more](/content/glossary/hash-tables/index.html)

H

## HellaSwag Reasoning Benchmark

A multiple-choice benchmark for commonsense natural language inference.

[Learn more](/content/glossary/hellaswag-reasoning-benchmark/index.html)

H

## Hellinger Distance

A metric for comparing probability distributions, related to Bhattacharyya coefficient.

[Learn more](/content/glossary/hellinger-distance/index.html)

H

## Helpfulness Metric

Scores how useful and goal-aligned an LLM response is for the user request.

[Learn more](/content/glossary/helpfulness-metric/index.html)

## Hijacking Excessive Agency Attack

An attack that manipulates an AI agent into acting beyond its intended permissions or tool scope.

[Learn more](/content/glossary/hijacking-excessive-agency-attack/index.html)

## Holdout Data

A train/test split portion reserved solely for unbiased model evaluation.

[Learn more](/content/glossary/holdout-data/index.html)

H

## HumanEval

A benchmark of programming problems scoring LLM-generated code via unit tests.

[Learn more](/content/glossary/humaneval/index.html)

H

## HumanEval Coding Benchmark

Functional-correctness benchmark for LLM code synthesis using unit tests.

[Learn more](/content/glossary/humaneval-coding-benchmark/index.html)

H

## Human-in-the-Loop Machine Learning

ML workflows that integrate human judgment for labeling, feedback, or oversight.

[Learn more](/content/glossary/human-in-the-loop-machine-learning/index.html)

H

## Hyperparameter Optimization

Systematic search for configuration values that maximize validation performance.

[Learn more](/content/glossary/hyperparameter-optimization/index.html)

H

## Hyperplane

A flat (n−1)-dimensional subspace used as a linear decision boundary in ML.

[Learn more](/content/glossary/hyperplane/index.html)

H

## Illegal Activities Harmful Content Attack

Red-team probe for prompts that push AI toward illegal or harmful criminal content.

[Learn more](/content/glossary/illegal-activities-harmful-content-attack/index.html)

## Image Data Augmentation

Artificially expanding image datasets with transforms to improve vision model generalization.

[Learn more](/content/glossary/image-data-augmentation/index.html)

I

## Image Data Collection

Gathering and organizing images for training computer vision models.

[Learn more](/content/glossary/image-data-collection/index.html)

I

## Imbalanced Data

Datasets where class frequencies are highly uneven, biasing naive classifiers.

[Learn more](/content/glossary/imbalanced-data/index.html)

I

## Impersonation Brand Damage Attack

Probe for AI impersonation of people or brands that could damage reputation.

[Learn more](/content/glossary/impersonation-brand-damage-attack/index.html)

## In-Context Learning

Adapting an LLM to a task via examples in the prompt rather than fine-tuning.

[Learn more](/content/glossary/in-context-learning/index.html)

I

## Independent and Identically Distributed Data IID

Data where samples are independent and share the same distribution.

[Learn more](/content/glossary/independent-and-identically-distributed-data-iid/index.html)

I

## Information Retrieval

Systems that find and rank documents relevant to a user query.

[Learn more](/content/glossary/information-retrieval/index.html)

I

## Instruction Tuning

Fine-tuning LLMs on instruction–response pairs to improve instruction following.

[Learn more](/content/glossary/instruction-tuning/index.html)

I

## Intelligent Document Processing IDP

OCR + NLP + ML pipelines that extract structured data from documents.

[Learn more](/content/glossary/intelligent-document-processing-idp/index.html)

I

## Intent Classification Metric

A metric that checks whether text is mapped to the correct intent label.

[Learn more](/content/glossary/intent-classification-metric/index.html)

## Intersection over Union (IoU)

Overlap ratio between predicted and ground-truth bounding boxes.

[Learn more](/content/glossary/intersection-over-union-iou/index.html)

I

## JSON Validation Metric

Automated checks that generated JSON is valid and includes required keys.

[Learn more](/content/glossary/json-validation-metric/index.html)

## Juries of Models Metric

Evaluating outputs with multiple LLM judges for more stable scores.

[Learn more](/content/glossary/juries-of-models-metric/index.html)

J

## Keras

A user-friendly high-level API for defining and training neural networks.

[Learn more](/content/glossary/keras/index.html)

K

## K-Nearest Neighbor KNN

A lazy supervised algorithm that classifies via the K closest labeled examples.

[Learn more](/content/glossary/k-nearest-neighbor-knn/index.html)

K

## KNN Models

Non-parametric models that predict from the K nearest training examples.

[Learn more](/content/glossary/knn-models/index.html)

K

## Kolmogorov-Smirnov Test

A nonparametric test comparing sample distributions via maximum CDF distance.

[Learn more](/content/glossary/kolmogorov-smirnov-test/index.html)

K

## KYC Process

Identity verification processes that deter fraud and meet financial regulations.

[Learn more](/content/glossary/kyc-process/index.html)

K

## LangChain

Open-source framework for building LLM chains, agents, and tools.

[Learn more](/content/glossary/langchain/index.html)

L

## Language Classification Metric

A metric that detects which natural language a text sample is written in.

[Learn more](/content/glossary/language-classification-metric/index.html)

## Large Action Models

Models that map natural-language goals into executable actions in digital environments.

[Learn more](/content/glossary/large-action-models/index.html)

L

## Learning Rate in Machine Learning

A hyperparameter controlling how much weights update on each training step.

[Learn more](/content/glossary/learning-rate-in-machine-learning/index.html)

L

## Learning-to-Rank

Supervised ML that optimizes the order of search or recommendation results.

[Learn more](/content/glossary/learning-to-rank/index.html)

L

## LegalBench Domain-Specific Benchmark

A 162-task benchmark covering six types of legal reasoning for AI evaluation.

[Learn more](/content/glossary/legalbench-domain-specific-benchmark/index.html)

L

## Liability Engagement Legal Risk Attack

Probe for AI agents making commitments that create legal or financial liability.

[Learn more](/content/glossary/liability-engagement-legal-risk-attack/index.html)

## LightGBM

A leaf-wise gradient boosting library optimized for speed on tabular data.

[Learn more](/content/glossary/lightgbm/index.html)

L

## Likert Framing Injection Attack

Jailbreak that frames harmful queries as Likert-scale research surveys.

[Learn more](/content/glossary/likert-framing-injection-attack/index.html)

## Linear Regression

A statistical method modeling a dependent variable as a linear combination of predictors.

[Learn more](/content/glossary/linear-regression/index.html)

L

## Link Injection Data Privacy Attack

Probe for AI generating phishing, malware, or other malicious links.

[Learn more](/content/glossary/link-injection-data-privacy-attack/index.html)

## Llama

Meta’s open LLM family used for research and many commercial applications.

[Learn more](/content/glossary/llama/index.html)

L

## LlamaIndex

Framework connecting datasets to LLMs for context-aware RAG apps.

[Learn more](/content/glossary/llamaindex/index.html)

L

## LLM Agents

Language-model systems that plan, call tools, and act toward user goals.

[Learn more](/content/glossary/llm-agents/index.html)

L

## LLM Alignment

Steering LLMs so behavior matches human values, intent, and safety policies.

[Learn more](/content/glossary/llm-alignment/index.html)

L

## LLM APIs

HTTP interfaces that let applications call large language models for text tasks.

[Learn more](/content/glossary/llm-apis/index.html)

L

## LLM-as-a-Service

Cloud-hosted access to large language models via APIs, without operating your own training or inference clusters.

[Learn more](/content/glossary/llm-as-a-service/index.html)

L

## LLM Benchmarks

Standardized task suites for comparing language model performance.

[Learn more](/content/glossary/llm-benchmarks/index.html)

L

## LLM Chatbot Evaluation

Systematic testing of chatbot quality, safety, and user experience.

[Learn more](/content/glossary/llm-chatbot-evaluation/index.html)

L

## LLM Cost

Token, infrastructure, and routing costs of operating large language models.

[Learn more](/content/glossary/llm-cost/index.html)

L

## LLM Debugger

Tools and practices for diagnosing LLM prompt and output failures.

[Learn more](/content/glossary/llm-debugger/index.html)

L

## LLM Deployment

Integrating LLMs into production with performance, safety, and ops controls.

[Learn more](/content/glossary/llm-deployment/index.html)

L

## LLM Distillation

Training a compact model to mimic a larger LLM’s outputs and skills.

[Learn more](/content/glossary/llm-distillation/index.html)

L

## LLM Embeddings

Dense vector representations of text used for search and retrieval.

[Learn more](/content/glossary/llm-embeddings/index.html)

L

## LLM Evaluation

Systematic measurement of LLM accuracy, safety, and usefulness.

[Learn more](/content/glossary/llm-evaluation/index.html)

L

## LLM Evaluation Framework

Structured protocol for assessing LLM performance, reliability, and ethics.

[Learn more](/content/glossary/llm-evaluation-framework/index.html)

L

## LLM Fine-Tuning

Further training a pre-trained LLM on specialized data for niche tasks.

[Learn more](/content/glossary/llm-fine-tuning/index.html)

L

## LLM Gateway

A proxy layer for routing, auth, and policy across LLM providers.

[Learn more](/content/glossary/llm-gateway/index.html)

L

## LLM Guardrails

Runtime rules and filters that keep LLM apps safe, compliant, and secure.

[Learn more](/content/glossary/llm-guardrails/index.html)

L

## LLM Hallucinations

When LLMs produce fluent but false or unsupported statements.

[Learn more](/content/glossary/llm-hallucinations/index.html)

L

## LLM Inference

Using a trained LLM to generate outputs on new, unseen inputs.

[Learn more](/content/glossary/llm-inference/index.html)

L

## LLM Interpretability

Methods that explain how LLMs turn inputs into outputs.

[Learn more](/content/glossary/llm-interpretability/index.html)

L

## LLM Jailbreaking

Bypassing LLM safety restrictions through adversarial prompts and multi-turn attacks.

[Learn more](/content/glossary/llm-jailbreaking/index.html)

L

## LLM Knowledge Base

Curated content used to ground LLM answers in trusted sources.

[Learn more](/content/glossary/llm-knowledge-base/index.html)

L

## LLM Knowledge Graph

Entity-relation graphs used to ground LLM answers with structured facts.

[Learn more](/content/glossary/llm-knowledge-graph/index.html)

L

## LLM Leaderboards

Public rankings of language models on shared benchmark suites.

[Learn more](/content/glossary/llm-leaderboards/index.html)

L

## LLM Observability

Monitoring prompts, traces, and quality signals for LLM applications.

[Learn more](/content/glossary/llm-observability/index.html)

L

## LLM Ontology

Concept hierarchies that structure domain knowledge for LLM reasoning.

[Learn more](/content/glossary/llm-ontology/index.html)

L

## LLMOps

Practices and tooling for deploying, monitoring, evaluating, and governing large language models in production systems.

[Learn more](/content/glossary/llmops/index.html)

L

## LLM Orchestration

Coordinating models, tools, and memory into multi-step AI workflows.

[Learn more](/content/glossary/llm-orchestration/index.html)

L

## LLM Output Parsing

Extracting structured, machine-usable data from free-form LLM responses for safe integration with tools and APIs.

[Learn more](/content/glossary/llm-output-parsing/index.html)

L

## LLM Overreliance

Excessive dependence on LLM outputs for decisions that still need human judgment, verification, and domain expertise.

[Learn more](/content/glossary/llm-overreliance/index.html)

L

## LLM Parameters

Model weights and decoding settings that control how large language models interpret prompts and generate outputs.

[Learn more](/content/glossary/llm-parameters/index.html)

L

## LLM Playground

An interactive environment for experimenting with LLM prompts, models, and parameters before production integration.

[Learn more](/content/glossary/llm-playground/index.html)

L

## LLM Product Development

Building products that embed large language models into user-facing workflows with clear value, evaluation, and iteration loops.

[Learn more](/content/glossary/llm-product-development/index.html)

L

## LLM Quantization

Reducing the numeric precision of model weights (and sometimes activations) to shrink size and speed inference with minimal quality loss.

[Learn more](/content/glossary/llm-quantization/index.html)

L

## LLM Red Teaming

Adversarial testing of large language models to find jailbreaks, data leaks, bias, and harmful outputs before deployment.

[Learn more](/content/glossary/llm-red-teaming/index.html)

L

## LLM Sleeper Agents

Language models with latent, trigger-activated behavior—often from fine-tuning or data poisoning—that can hide backdoors until conditions fire.

[Learn more](/content/glossary/llm-sleeper-agents/index.html)

L

## LLM Stack Layers

The layered toolchain for LLM apps: data, model, orchestration, serving, evaluation, and interface components.

[Learn more](/content/glossary/llm-stack-layers/index.html)

L

## LLM Summarization

Using large language models to condense long documents into shorter text while preserving key information.

[Learn more](/content/glossary/llm-summarization/index.html)

L

## LLM Testing

Systematic evaluation of large language models for accuracy, robustness, safety, and reliability before production use.

[Learn more](/content/glossary/llm-testing/index.html)

L

## LLM Toxicity

Hate, harassment, or abusive content produced by large language models, often learned from unfiltered training data.

[Learn more](/content/glossary/llm-toxicity/index.html)

L

## LLM Tracing

Recording prompts, responses, tool calls, latency, and costs across LLM application runs for debugging and observability.

[Learn more](/content/glossary/llm-tracing/index.html)

L

## Locally Interpretable Model-Agnostic Explanations LIME

A model-agnostic method that explains individual predictions with a simple local surrogate model around the instance.

[Learn more](/content/glossary/locally-interpretable-model-agnostic-explanations-lime/index.html)

L

## Logistic Regression

A classification method that models the probability of categorical outcomes using a logistic (sigmoid) function.

[Learn more](/content/glossary/logistic-regression/index.html)

L

## Long Short Term Memory LSTM

A recurrent neural network architecture that retains long-range sequence information via gated memory cells.

[Learn more](/content/glossary/long-short-term-memory-lstm/index.html)

L

## Low Rank Adaptation of Large Language Models

A parameter-efficient fine-tuning method that adapts LLMs with small low-rank matrices instead of full weight updates.

[Learn more](/content/glossary/low-rank-adaptation-of-large-language-models/index.html)

L

## Machine Learning

A field of AI where algorithms learn patterns from data to make predictions or decisions without being fully hard-coded.

[Learn more](/content/glossary/machine-learning/index.html)

M

## Machine Learning Algorithm

A computational method that learns patterns from data to predict or decide on unseen examples.

[Learn more](/content/glossary/machine-learning-algorithm/index.html)

M

## Machine Learning as a Service (MLaaS)

Cloud services that provide tools and APIs to build, train, and deploy machine learning models without owning the full stack.

[Learn more](/content/glossary/machine-learning-as-a-service-mlaas/index.html)

M

## Machine Learning Bias

Systematic skew in ML predictions that unfairly favors or disadvantages groups, often from biased training data or objectives.

[Learn more](/content/glossary/machine-learning-bias/index.html)

M

## Machine Learning Checkpointing

Periodically saving model and optimizer state during training to enable recovery, comparison, and rollback.

[Learn more](/content/glossary/machine-learning-checkpointing/index.html)

M

## Machine Learning Deployment

Integrating trained ML models into production applications and APIs so they serve live predictions.

[Learn more](/content/glossary/machine-learning-deployment/index.html)

M

## Machine Learning Inference

Executing a trained model on new data to produce predictions or generations outside the training loop.

[Learn more](/content/glossary/machine-learning-inference/index.html)

M

## Machine Learning in Software Testing

Using ML to generate, prioritize, or analyze software tests for better coverage and faster defect detection.

[Learn more](/content/glossary/machine-learning-in-software-testing/index.html)

M

## Machine Learning Lifecycle

End-to-end stages of an ML project: problem framing, data, training, validation, deployment, and monitoring.

[Learn more](/content/glossary/machine-learning-lifecycle/index.html)

M

## Machine Learning Model Accuracy

The proportion of correct predictions; often paired with precision, recall, and slice-aware metrics for decisions.

[Learn more](/content/glossary/machine-learning-model-accuracy/index.html)

M

## Machine Learning Model Evaluation

Assessing ML model effectiveness with held-out data and metrics like accuracy, AUC, F1-score, and calibration.

[Learn more](/content/glossary/machine-learning-model-evaluation/index.html)

M

## Machine Learning Orchestration

Automating and coordinating ML pipeline stages—from data prep to deployment and monitoring—at scale.

[Learn more](/content/glossary/machine-learning-orchestration/index.html)

M

## Machine Learning Pipeline

A structured workflow from data preparation through training, evaluation, deployment, and monitoring of ML models.

[Learn more](/content/glossary/machine-learning-pipeline/index.html)

M

## Machine Learning Workflows

Ordered steps for delivering ML solutions: data collection, preparation, training, evaluation, and iteration.

[Learn more](/content/glossary/machine-learning-workflows/index.html)

M

## MAP (Mean Average Precision) Metric

A ranking metric that averages precision across queries, widely used for search and retrieval evaluation.

[Learn more](/content/glossary/map-mean-average-precision-metric/index.html)

M

## Math Framing Injection Attack

A jailbreak pattern that hides harmful intents inside mathematical or academic framing to bypass AI safety filters.

[Learn more](/content/glossary/math-framing-injection-attack/index.html)

## MATH Math Benchmark

A benchmark of competition-style math problems used to evaluate LLM reasoning from elementary through high-school topics.

[Learn more](/content/glossary/math-math-benchmark/index.html)

M

## MBPP Coding Benchmark

A Python coding benchmark with hundreds of problems assessing correctness and core programming skill in generated solutions.

[Learn more](/content/glossary/mbpp-coding-benchmark/index.html)

M

## Mean Absolute Error

Average absolute gap between predictions and true values; a common regression accuracy metric.

[Learn more](/content/glossary/mean-absolute-error/index.html)

M

## Mean Absolute Percentage Error

Average absolute percentage difference between predicted and actual values; common in forecasting evaluation.

[Learn more](/content/glossary/mean-absolute-percentage-error/index.html)

M

## Mean Squared Error (MSE)

Average of squared prediction errors; a common regression loss and accuracy metric.

[Learn more](/content/glossary/mean-squared-error-mse/index.html)

M

## Memory-Augmented Neural Networks

Neural networks extended with external memory for longer-term storage and retrieval during complex tasks.

[Learn more](/content/glossary/memory-augmented-neural-networks/index.html)

M

## Meta Learning

A paradigm where algorithms learn how to learn, enabling fast adaptation to new tasks from limited examples.

[Learn more](/content/glossary/meta-learning/index.html)

M

## METEOR Score

A translation evaluation metric that aligns outputs to references using stems and synonyms, balancing precision and recall.

[Learn more](/content/glossary/meteor-score/index.html)

M

## Micro-Models

Compact models focused on narrow tasks or data slices within a larger modeling system.

[Learn more](/content/glossary/micro-models/index.html)

M

## Misinformation & Disinformation Harmful Content Attack

An adversarial probe that tries to make AI agents generate or amplify false, misleading, or harmful information.

[Learn more](/content/glossary/misinformation-disinformation-harmful-content-attack/index.html)

## Missing Values in Time Series

Absent observations in sequential datasets that require removal, imputation, or specialized models to avoid biased analysis.

[Learn more](/content/glossary/missing-values-in-time-series/index.html)

M

## Mixture of Experts

An architecture that routes each input to a small subset of expert subnetworks for scalable model capacity.

[Learn more](/content/glossary/mixture-of-experts/index.html)

M

## ML Architecture

The blueprint for how data, models, serving, and monitoring components connect in an ML system.

[Learn more](/content/glossary/ml-architecture/index.html)

M

## ML-Based Scoring Metric

Cost-effective text evaluation with pre-trained scorers instead of LLM-as-judge.

[Learn more](/content/glossary/ml-based-scoring-metric/index.html)

M

## ML Diagnostics

Systematic identification and remediation of performance, bias, and robustness issues in ML models.

[Learn more](/content/glossary/ml-diagnostics/index.html)

M

## ML Infrastructure

Compute, storage, and tooling foundations for building, training, and deploying machine learning models.

[Learn more](/content/glossary/ml-infrastructure/index.html)

M

## ML Interpretability

Understanding and describing how ML models make predictions, beyond raw accuracy scores.

[Learn more](/content/glossary/ml-interpretability/index.html)

M

## ML Model Cards

Standardized documentation of an ML model’s use, data, metrics, biases, and limitations.

[Learn more](/content/glossary/ml-model-cards/index.html)

M

## ML Model Management

Systems and practices for versioning, approving, deploying, and monitoring machine learning models.

[Learn more](/content/glossary/ml-model-management/index.html)

M

## ML Model Validation

Confirming that models meet quantitative and qualitative requirements on intended data before deployment.

[Learn more](/content/glossary/ml-model-validation/index.html)

M

## MLOps

Practices that connect ML development and operations for scalable, governed deployments.

[Learn more](/content/glossary/mlops/index.html)

M

## MLOps for Generative AI

CI/CD, monitoring, and governance adapted for generative AI systems and agents.

[Learn more](/content/glossary/mlops-for-generative-ai/index.html)

M

## MLOps Framework

A structured methodology for automating and governing the ML lifecycle end to end.

[Learn more](/content/glossary/mlops-framework/index.html)

M

## MLOps Monitoring

Tracking live ML performance, drift, and feedback so production models stay trustworthy.

[Learn more](/content/glossary/mlops-monitoring/index.html)

M

## ML Performance Tracing

Monitoring model predictions and operational metrics across the lifecycle to find bottlenecks and quality drift.

[Learn more](/content/glossary/ml-performance-tracing/index.html)

M

## ML Scalability

Ability of an ML system to process more data or traffic without unacceptable loss of performance or precision.

[Learn more](/content/glossary/ml-scalability/index.html)

M

## ML Stack

The collection of software tools and frameworks used to develop, deploy, and analyze machine learning solutions.

[Learn more](/content/glossary/ml-stack/index.html)

M

## MMLU benchmark

Massive Multitask Language Understanding: a broad academic benchmark for LLM capability.

[Learn more](/content/glossary/mmlu-benchmark/index.html)

M

## MMLU Reasoning Benchmark

Multiple-choice academic questions used to probe an LLM knowledge and reasoning breadth.

[Learn more](/content/glossary/mmlu-reasoning-benchmark/index.html)

M

## Model-Based Machine Learning (MBML)

Tailoring probabilistic models to problems via factor graphs and Bayesian inference.

[Learn more](/content/glossary/model-based-machine-learning-mbml/index.html)

M

## Model Behavior

How a model predicts and behaves under real inputs, shaped by data, structure, and training.

[Learn more](/content/glossary/model-behavior/index.html)

M

## Model Calibration

Adjusting predicted probabilities so they match observed outcome frequencies.

[Learn more](/content/glossary/model-calibration/index.html)

M

## Model Distillation

Compressing a large teacher model into a smaller student that retains most of its behavior.

[Learn more](/content/glossary/model-distillation/index.html)

M

## Model Drift

Performance loss when live data or conditions diverge from what the model was trained on.

[Learn more](/content/glossary/model-drift/index.html)

M

## Model-Driven Architecture

Software development centered on platform-independent models and transformations.

[Learn more](/content/glossary/model-driven-architecture/index.html)

M

## Model Explainability

Methods that make ML decisions understandable to stakeholders and auditors.

[Learn more](/content/glossary/model-explainability/index.html)

M

## Model Fairness

Ensuring ML predictions avoid unjustified disparate impact across groups.

[Learn more](/content/glossary/model-fairness/index.html)

M

## Model Merging

Combining multiple models weights or adapters into one stronger or more specialized model.

[Learn more](/content/glossary/model-merging/index.html)

M

## Model Monitoring

Real-time observation of ML performance to detect failures and meet business requirements.

[Learn more](/content/glossary/model-monitoring/index.html)

M

## Model Observability

Inspecting ML inputs, outputs, and internals to debug and optimize production systems.

[Learn more](/content/glossary/model-observability/index.html)

M

## Model Parameters

Learned internal coefficients that determine how a model transforms inputs into predictions.

[Learn more](/content/glossary/model-parameters/index.html)

M

## Model Registry

Centralized versioning and lifecycle management for production-ready ML models.

[Learn more](/content/glossary/model-registry/index.html)

M

## Model Retraining

Updating a trained model with new data to improve performance or adapt to change.

[Learn more](/content/glossary/model-retraining/index.html)

M

## Model Robustness

A model's ability to stay accurate under noise, shifts, and adversarial conditions.

[Learn more](/content/glossary/model-robustness/index.html)

M

## Model Selection

Choosing the best model among candidates using validation and resampling techniques.

[Learn more](/content/glossary/model-selection/index.html)

M

## Model Tuning

Adjusting hyperparameters to optimize a model's learning process and accuracy.

[Learn more](/content/glossary/model-tuning/index.html)

M

## MRR (Mean Reciprocal Rank) Metric

Ranking metric focused on how soon the first correct answer appears.

[Learn more](/content/glossary/mrr-mean-reciprocal-rank-metric/index.html)

M

## MT-Bench

A multi-turn benchmark for conversational LLM quality and coherence.

[Learn more](/content/glossary/mt-bench/index.html)

M

## MT-Bench Conversation Benchmark

Benchmark for how well LLMs stay coherent across conversation turns.

[Learn more](/content/glossary/mt-bench-conversation-benchmark/index.html)

M

## MTEB

Massive Text Embedding Benchmark for comparing embedding model quality.

[Learn more](/content/glossary/mteb/index.html)

M

## Multi-Class Classification

Predicting one class among three or more possible labels for each instance.

[Learn more](/content/glossary/multi-class-classification/index.html)

M

## Multilayer Perceptron (MLP)

Feedforward neural net with one or more hidden layers for classification and regression.

[Learn more](/content/glossary/multilayer-perceptron-mlp/index.html)

M

## Multilingual LLM

Large language models trained to understand and generate text in many languages.

[Learn more](/content/glossary/multilingual-llm/index.html)

M

## MultiMedQA Domain-Specific Benchmark

Medical QA benchmark combining datasets to score factuality, harm, and bias.

[Learn more](/content/glossary/multimedqa-domain-specific-benchmark/index.html)

M

## Naive Bayes Model

Fast probabilistic classifier using Bayes theorem with independent-feature assumptions.

[Learn more](/content/glossary/naive-bayes-model/index.html)

N

## Natural Language Search

Search that understands user intent from free-form language, not just keyword match.

[Learn more](/content/glossary/natural-language-search/index.html)

N

## Natural Language Understanding

AI methods that let machines interpret meaning and intent in human language.

[Learn more](/content/glossary/natural-language-understanding/index.html)

N

## NDCG (Normalized Discounted Cumulative Gain) Metric

Ranking metric that rewards relevant items higher up the list, normalized to an ideal order.

[Learn more](/content/glossary/ndcg-normalized-discounted-cumulative-gain-metric/index.html)

N

## Neural Networks

Layered models of artificial neurons used for learning patterns from data.

[Learn more](/content/glossary/neural-networks/index.html)

N

## Neural Network Tuning

Optimizing neural net hyperparameters like depth, learning rate, and batch size.

[Learn more](/content/glossary/neural-network-tuning/index.html)

N

## No-code Low-code ML

Platforms that build ML apps and workflows with visual tools instead of heavy coding.

[Learn more](/content/glossary/no-code-low-code-ml/index.html)

N

## Noise in Machine Learning

Unwanted data variation that can mislead training and degrade generalization.

[Learn more](/content/glossary/noise-in-machine-learning/index.html)

N

## Noisy Image

Images with unwanted brightness or color variation that may need denoising.

[Learn more](/content/glossary/noisy-image/index.html)

N

## Normalization in Machine Learning

Rescaling features to a common range (e.g., min-max) for stable model training.

[Learn more](/content/glossary/normalization-in-machine-learning/index.html)

N

## Normalized Discounted Cumulative Gain

Ranking effectiveness metric balancing relevance and position, normalized to ideal order.

[Learn more](/content/glossary/normalized-discounted-cumulative-gain/index.html)

N

## Nvidia NIM

NVIDIA's inference microservices for deploying optimized AI models at scale.

[Learn more](/content/glossary/nvidia-nim/index.html)

N

## Observation ML

A single data instance (feature vector) analyzed or predicted by a learning algorithm.

[Learn more](/content/glossary/observation-ml/index.html)

O

## One-Hot Encoding

Representing categories as binary columns so models can process categorical features.

[Learn more](/content/glossary/one-hot-encoding/index.html)

O

## Online Machine Learning

Incremental learning from streaming data so models adapt without full batch retrains.

[Learn more](/content/glossary/online-machine-learning/index.html)

O

## Open-Source Machine Learning Monitoring

Community tools for supervising live ML models—drift, performance, and alerts.

[Learn more](/content/glossary/open-source-machine-learning-monitoring/index.html)

O

## Out-of-distribution

Inputs that differ substantially from training data, often causing unreliable predictions.

[Learn more](/content/glossary/out-of-distribution/index.html)

O

## Overfitting in Machine Learning

Learning training noise so closely that performance on new data drops.

[Learn more](/content/glossary/overfitting-in-machine-learning/index.html)

O

## OWASP AI Red Teaming Solutions Landscape Taxonomy

OWASP grid of AI red-teaming capabilities mapped to lifecycle stages from Plan and Test through Operate and Govern.

[Learn more](/content/glossary/owasp-ai-red-teaming-solutions-landscape-taxonomy/index.html)

O

## OWASP ASI01 Agent Goal Hijack

When email, PDFs, or tool output rewrite an agent objective so the workflow chases the wrong goal.

[Learn more](/content/glossary/owasp-asi01-agent-goal-hijack-2e3z3/index.html)

A

## OWASP ASI02 Tool Misuse and Exploitation

When agents misuse real tools and permissions after injection or misalignment.

[Learn more](/content/glossary/owasp-asi02-tool-misuse-and-exploitation-nu0lr/index.html)

T

## OWASP ASI03 Identity and Privilege Abuse

Delegation chains and inherited credentials that escalate access without re-checking the user.

[Learn more](/content/glossary/owasp-asi03-identity-and-privilege-abuse-mhogs/index.html)

I

## OWASP ASI04 Agentic Supply Chain Vulnerabilities

Runtime-loaded MCP servers, prompt packs, and agent personas you did not fully trust.

[Learn more](/content/glossary/owasp-asi04-agentic-supply-chain-vulnerabilities-zbanx/index.html)

A

## OWASP ASI05 Unexpected Code Execution

Coding agents and vibe-coding tools running shell commands nobody reviewed.

[Learn more](/content/glossary/owasp-asi05-unexpected-code-execution-jpdmy/index.html)

U

## OWASP ASI06 Memory and Context Poisoning

Bad facts that persist across sessions and compound every future task.

[Learn more](/content/glossary/owasp-asi06-memory-and-context-poisoning-io5nt/index.html)

M

## OWASP ASI07 Insecure Inter-Agent Communication

ASI07: weak auth, spoofed peers, and unencrypted A2A/MCP traffic between agents.

[Learn more](/content/glossary/owasp-asi07-insecure-inter-agent-communication-95svy/index.html)

I

## OWASP ASI08 Cascading Failures

ASI08: one poisoned agent output cascading through a multi-agent workflow before humans can stop it.

[Learn more](/content/glossary/owasp-asi08-cascading-failures-itpv6/index.html)

C

## OWASP ASI09 Human-Agent Trust Exploitation

ASI09: agents using urgency or authority cues to push humans into unsafe approvals.

[Learn more](/content/glossary/owasp-asi09-human-agent-trust-exploitation-i2o37/index.html)

H

## OWASP ASI10 Rogue Agents

ASI10: misaligned or self-persisting agents that keep acting after they should be revoked.

[Learn more](/content/glossary/owasp-asi10-rogue-agents-azfzg/index.html)

R

## OWASP LLM01 (2023) Prompt Injection

The 2023/24 OWASP #1 LLM risk: crafted inputs that override developer instructions.

[Learn more](/content/glossary/owasp-llm01-2023-prompt-injection-y3vck/index.html)

P

## OWASP LLM01 (2025) Prompt Injection

OWASP #1 LLM risk: direct, indirect, and automated prompt injection that overrides instructions.

[Learn more](/content/glossary/owasp-llm01-2025-prompt-injection-34i2g/index.html)

P

## OWASP LLM02 (2023) Insecure Output Handling

2023 OWASP risk: apps trusting LLM output into browsers, shells, or APIs without sanitization.

[Learn more](/content/glossary/owasp-llm02-2023-insecure-output-handling-mrz6j/index.html)

I

## OWASP LLM02 (2025) Sensitive Information Disclosure

When LLMs leak PII, credentials, or internal context through chat, tools, or cross-session memory.

[Learn more](/content/glossary/owasp-llm02-2025-sensitive-information-disclosure-lu0iy/index.html)

S

## OWASP LLM03 (2023) Training Data Poisoning

2023 OWASP entry for poisoned training or fine-tuning data that embeds backdoors or bias.

[Learn more](/content/glossary/owasp-llm03-2023-training-data-poisoning-zmpqq/index.html)

T

## OWASP LLM03 (2025) Supply Chain

LLM supply-chain risk from models, plugins, datasets, and MCP servers you did not build.

[Learn more](/content/glossary/owasp-llm03-2025-supply-chain-zk8qc/index.html)

S

## OWASP LLM04 (2023) Model Denial of Service

2023 OWASP risk: resource exhaustion that degrades LLM availability for legitimate users.

[Learn more](/content/glossary/owasp-llm04-2023-model-denial-of-service-ykfd7/index.html)

M

## OWASP LLM04 (2025) Data and Model Poisoning

Poisoned training, fine-tuning, RAG, or memory that steers trusted LLM outputs wrong.

[Learn more](/content/glossary/owasp-llm04-2025-data-and-model-poisoning-be3ng/index.html)

## OWASP LLM05 (2023) Supply Chain Vulnerabilities

2023 OWASP risk from untrusted models, datasets, plugins, and weights in the LLM stack.

[Learn more](/content/glossary/owasp-llm05-2023-supply-chain-vulnerabilities-ac2n9/index.html)

S

## OWASP LLM05 (2025) Improper Output Handling

When trusted LLM text becomes XSS, SQL injection, or shell execution in downstream systems.

[Learn more](/content/glossary/owasp-llm05-2025-improper-output-handling-s002j/index.html)

I

## OWASP LLM06 (2023) Sensitive Information Disclosure

2023 OWASP entry for LLMs revealing confidential data in responses or context.

[Learn more](/content/glossary/owasp-llm06-2023-sensitive-information-disclosure-ffk79/index.html)

S

## OWASP LLM06 (2025) Excessive Agency

When an LLM can call APIs, write data, or send messages without enough human checks.

[Learn more](/content/glossary/owasp-llm06-2025-excessive-agency-oprdx/index.html)

E

## OWASP LLM07 (2023) Insecure Plugin Design

2023 OWASP risk from poorly designed LLM plugins that escalate privilege or trust input.

[Learn more](/content/glossary/owasp-llm07-2023-insecure-plugin-design-icemc/index.html)

I

## OWASP LLM07 (2025) System Prompt Leakage

When crafted prompts extract system instructions, guardrails, or tool schemas.

[Learn more](/content/glossary/owasp-llm07-2025-system-prompt-leakage-3imtc/index.html)

S

## OWASP LLM08 (2023) Excessive Agency

2023 OWASP risk: LLM agents granted more autonomy and tools than safety controls justify.

[Learn more](/content/glossary/owasp-llm08-2023-excessive-agency-a9x49/index.html)

E

## OWASP LLM08 (2025) Vector and Embedding Weaknesses

RAG attack surface: poisoned chunks, weak vector DB access, and manipulated similarity.

[Learn more](/content/glossary/owasp-llm08-2025-vector-and-embedding-weaknesses-fc7tr/index.html)

V

## OWASP LLM09 (2023) Overreliance

2023 OWASP risk: people and systems trusting fluent LLM answers without verification.

[Learn more](/content/glossary/owasp-llm09-2023-overreliance-nwabw/index.html)

O

## OWASP LLM09 (2025) Misinformation

Fluent, authoritative LLM falsehoods that drive bad decisions in finance, legal, or medical use.

[Learn more](/content/glossary/owasp-llm09-2025-misinformation-k06ah/index.html)

M

## OWASP LLM10 (2023) Model Theft

2023 OWASP risk of stealing model weights or approximating behavior through queries.

[Learn more](/content/glossary/owasp-llm10-2023-model-theft-dn0vm/index.html)

M

## OWASP LLM10 (2025) Unbounded Consumption

Runaway tokens, tool loops, and inference spend that drain budget without a hard crash.

[Learn more](/content/glossary/owasp-llm10-2025-unbounded-consumption-prnqd/index.html)

U

## Pandas and NumPy

Pandas and NumPy are core Python libraries for tabular analysis and fast numerical array computation.

[Learn more](/content/glossary/pandas-and-numpy/index.html)

P

## Parameter-Efficient Fine-Tuning

PEFT adapts LLMs by updating a small parameter set (e.g. LoRA) instead of full fine-tuning.

[Learn more](/content/glossary/parameter-efficient-fine-tuning/index.html)

P

## Parameter-Efficient Fine-Tuning (Prefix-Tuning)

Prefix-tuning learns continuous prefix vectors while freezing most base LLM weights.

[Learn more](/content/glossary/parameter-efficient-fine-tuning-prefix-tuning/index.html)

P

## Pascal

PASCAL VOC is a classic object detection and segmentation benchmark with standardized bounding boxes.

[Learn more](/content/glossary/pascal/index.html)

P

## Pattern Matching

Pattern matching finds exact or regex-defined patterns in sequences of data or tokens.

[Learn more](/content/glossary/pattern-matching/index.html)

P

## Pattern Recognition

Pattern recognition identifies regularities in data so systems can classify or detect structures.

[Learn more](/content/glossary/pattern-recognition/index.html)

P

## Permutation Importance

Permutation importance estimates feature influence by measuring performance drop after shuffling a feature.

[Learn more](/content/glossary/permutation-importance/index.html)

P

## Phare Safety Benchmark

Phare is a multilingual LLM safety benchmark for probing harmful content and refusal behavior.

[Learn more](/content/glossary/phare-safety-benchmark/index.html)

P

## PII Detection Metric

A PII detection metric scores whether model outputs expose personally identifiable information.

[Learn more](/content/glossary/pii-detection-metric/index.html)

## PII Leak Data Privacy Attack

A PII leak attack probes whether an LLM reveals personally identifiable information it should withhold.

[Learn more](/content/glossary/pii-leak-data-privacy-attack/index.html)

## Politeness Metric

A politeness metric scores whether model responses stay courteous and appropriately toned.

[Learn more](/content/glossary/politeness-metric/index.html)

## Politics Topic Harmful Content Attack

Safety probe checking whether an LLM refuses disallowed political harm content.

[Learn more](/content/glossary/politics-topic-harmful-content-attack/index.html)

## Pooling Layers in CNN

Pooling layers downsample CNN feature maps to reduce spatial size and keep salient activations.

[Learn more](/content/glossary/pooling-layers-in-cnn/index.html)

P

## Population Stability Index

PSI measures how much a feature distribution shifted between a reference and current population.

[Learn more](/content/glossary/population-stability-index/index.html)

P

## Positional Encoding

Positional encoding injects sequence-order information into transformer inputs that lack position.

[Learn more](/content/glossary/positional-encoding/index.html)

P

## Precision in Machine Learning

Precision is the fraction of predicted positives that are actually positive.

[Learn more](/content/glossary/precision-in-machine-learning/index.html)

P

## Precision Metric

A precision metric scores how many of the systems claimed positives are correct.

[Learn more](/content/glossary/precision-metric/index.html)

## Precision-Recall Area Under the Curve (PR AUC)

PR AUC summarizes the precision-recall curve, especially useful on imbalanced classification problems.

[Learn more](/content/glossary/precision-recall-area-under-the-curve-pr-auc/index.html)

P

## Predictive Model Validation

Predictive model validation estimates generalization via holdout sets, cross-validation, or backtests.

[Learn more](/content/glossary/predictive-model-validation/index.html)

P

## Preprocessing

Preprocessing transforms raw inputs into features models can learn from.

[Learn more](/content/glossary/preprocessing/index.html)

P

## Pre-trained Transformer

A pre-trained transformer is trained on large corpora then adapted downstream via fine-tuning or prompting.

[Learn more](/content/glossary/pre-trained-transformer/index.html)

P

## Principal Component Analysis (PCA)

PCA projects data onto orthogonal components that capture maximal variance.

[Learn more](/content/glossary/principal-component-analysis-pca/index.html)

P

## Probabilistic Classification

Probabilistic classification predicts class probabilities, not only hard labels.

[Learn more](/content/glossary/probabilistic-classification/index.html)

P

## Programmatic Validation Metric

A programmatic validation metric scores outputs with deterministic rules such as schemas or regex.

[Learn more](/content/glossary/programmatic-validation-metric/index.html)

## Prompt Alignment Metric

A prompt alignment metric scores how well a response follows user instructions and constraints.

[Learn more](/content/glossary/prompt-alignment-metric/index.html)

## Prompt Chaining

Prompt chaining sequences multiple LLM calls so each steps output feeds the next stage.

[Learn more](/content/glossary/prompt-chaining/index.html)

P

## Prompt Engineering

Prompt engineering designs instructions and context that steer LLM behavior toward reliable outcomes.

[Learn more](/content/glossary/prompt-engineering/index.html)

P

## Prompt Extraction Internal Information Attack

A prompt extraction attack tries to coerce an LLM into revealing its system prompt or hidden instructions.

[Learn more](/content/glossary/prompt-extraction-internal-information-attack/index.html)

## Prompt Injection

Prompt injection occurs when untrusted text overrides developer instructions and steers model behavior.

[Learn more](/content/glossary/prompt-injection/index.html)

P

## Prompt Playground

A prompt playground is an interactive UI for trying prompts, models, and parameters during development.

[Learn more](/content/glossary/prompt-playground/index.html)

P

## Prototype Model

A prototype model is an early proof-of-concept trained to validate feasibility before production hardening.

[Learn more](/content/glossary/prototype-model/index.html)

P

## Purple Teaming Capabilities (OWASP Taxonomy)

OWASP purple teaming: map red findings to blue controls and measure what actually works.

[Learn more](/content/glossary/purple-teaming-capabilities-owasp-taxonomy/index.html)

P

## PyTorch

PyTorch is an open-source deep learning framework with dynamic graphs and pythonic tensor APIs.

[Learn more](/content/glossary/pytorch/index.html)

P

## RAG Architecture

System design that retrieves external context and feeds it into a generative model for grounded answers.

[Learn more](/content/glossary/rag-architecture/index.html)

R

## RAGAS

An evaluation framework for measuring RAG pipeline quality and reliability.

[Learn more](/content/glossary/ragas/index.html)

R

## RAG as a Service

Cloud-managed RAG that packages retrieval, indexing, and generation for applications.

[Learn more](/content/glossary/rag-as-a-service/index.html)

R

## RAGET Complex Question Hallucination Attack

Uses complex questions derived from your knowledge base to test RAG hallucination resistance.

[Learn more](/content/glossary/raget-complex-question-hallucination-attack/index.html)

## RAGET Simple Question Hallucination Attack

Generates simple factual questions from your knowledge base to evaluate RAG answer accuracy.

[Learn more](/content/glossary/raget-simple-question-hallucination-attack/index.html)

## RAGET Situational Question Hallucination Attack

Tests RAG hallucinations using situational and user-specific question framing drawn from your docs.

[Learn more](/content/glossary/raget-situational-question-hallucination-attack/index.html)

## RAG Evaluation

How to assess RAG systems across retrieval quality and generated answer faithfulness.

[Learn more](/content/glossary/rag-evaluation/index.html)

R

## RAG Hallucinations

When retrieval-augmented models still produce false or unsupported claims that sound credible.

[Learn more](/content/glossary/rag-hallucinations/index.html)

R

## Random Forest

Ensemble of decision trees using bagging and random feature selection.

[Learn more](/content/glossary/random-forest/index.html)

R

## Random Initialization

Initializing network weights randomly near zero to break symmetry for gradient descent.

[Learn more](/content/glossary/a-technique-to-initialize-neural-network-weights-with-random-values-close-to-zero-ensuring-diverse-neuron-outputs-and-aiding-efficient-gradient-descent/index.html)

R

## Readability Assessment Metric

Scores how easy AI-generated text is for a target audience to read.

[Learn more](/content/glossary/readability-assessment-metric/index.html)

## Reasoning Engine

Component that performs structured inference or planning to produce conclusions or actions.

[Learn more](/content/glossary/reasoning-engine/index.html)

R

## Recall in Machine Learning

Fraction of true positives identified among all actual positive cases.

[Learn more](/content/glossary/recall-in-machine-learning/index.html)

R

## Recall Metric

The share of true positives found among all actual positives.

[Learn more](/content/glossary/recall-metric/index.html)

## Recall-Oriented Understudy for Gisting Evaluation (ROUGE)

N-gram recall metrics for comparing generated text to reference summaries.

[Learn more](/content/glossary/recall-oriented-understudy-for-gisting-evaluation-rouge/index.html)

R

## Receiver Operating Characteristic (ROC) Curve

Plot of TPR vs FPR used to assess binary classifier discrimination.

[Learn more](/content/glossary/receiver-operating-characteristic-roc-curve/index.html)

R

## Rectified Linear Unit (ReLU)

Activation that zeros negatives and passes positives unchanged.

[Learn more](/content/glossary/rectified-linear-unit-relu/index.html)

R

## Recurrent Neural Networks

Neural nets with temporal state for sequential data.

[Learn more](/content/glossary/recurrent-neural-networks/index.html)

R

## Red Teaming Capabilities (OWASP Taxonomy)

OWASP taxonomy red-teaming capabilities: adversarial probes, multi-turn attacks, and continuous offensive testing for AI systems.

[Learn more](/content/glossary/red-teaming-capabilities-owasp-taxonomy/index.html)

R

## Reference Distribution

Baseline distribution against which samples or model outputs are compared.

[Learn more](/content/glossary/reference-distribution/index.html)

R

## Regex Pattern Matching Metric

Deterministic checks that match or reject outputs using regular expressions.

[Learn more](/content/glossary/regex-pattern-matching-metric/index.html)

## Regression

Supervised learning that estimates continuous outcomes.

[Learn more](/content/glossary/regression/index.html)

R

## Regression Algorithms

Family of methods that map features to continuous targets.

[Learn more](/content/glossary/regression-algorithms/index.html)

R

## Regularization Algorithms

Techniques that penalize complexity to reduce overfitting.

[Learn more](/content/glossary/regularization-algorithms/index.html)

R

## Regularization in Machine Learning

Constraining model complexity so predictors generalize to new data.

[Learn more](/content/glossary/regularization-in-machine-learning/index.html)

R

## Reinforcement learning

Learning optimal actions via rewards and penalties over time.

[Learn more](/content/glossary/reinforcement-learning/index.html)

R

## Reinforcement Learning from AI Feedback (RLAIF)

Aligning models with preference signals produced by AI judges.

[Learn more](/content/glossary/reinforcement-learning-from-ai-feedback-rlaif/index.html)

R

## Relevance Metric

Scores how well results or answers match the user's information need.

[Learn more](/content/glossary/relevance-metric/index.html)

## Religion Topic Harmful Content Attack

Adversarial prompts that provoke biased or harmful statements about religion to test fairness.

[Learn more](/content/glossary/religion-topic-harmful-content-attack/index.html)

## Reproducible AI

Practices that make AI experiments and evaluations repeatable across runs and teams.

[Learn more](/content/glossary/reproducible-ai/index.html)

R

## ResNet

CNN architecture with residual connections enabling deeper training.

[Learn more](/content/glossary/resnet/index.html)

R

## Responsible AI

Principles and practices for building AI that is safe, fair, transparent, and accountable.

[Learn more](/content/glossary/responsible-ai/index.html)

R

## Retrieval-augmented Generation (RAG)

LLM pattern that retrieves external knowledge before generating a response.

[Learn more](/content/glossary/retrieval-augmented-generation-rag/index.html)

R

## Ridge Regression

Linear model with squared-coefficient penalty for stability.

[Learn more](/content/glossary/ridge-regression/index.html)

R

## RMSprop

Optimizer that scales steps by recent gradient magnitude.

[Learn more](/content/glossary/rmsprop/index.html)

R

## Robotic Process Automation (RPA)

Automating structured business processes with software robots.

[Learn more](/content/glossary/robotic-process-automation-rpa/index.html)

R

## ROC-AUC Metric

Area under the ROC curve; measures ranking quality independent of a single threshold.

[Learn more](/content/glossary/roc-auc-metric/index.html)

R

## Root-Cause Analysis

Structured investigation to find underlying causes of failures.

[Learn more](/content/glossary/root-cause-analysis/index.html)

R

## Root Mean Square Error (RMSE)

Square root of mean squared error between predictions and targets.

[Learn more](/content/glossary/root-mean-square-error-rmse/index.html)

R

## Rotating Proxy

Proxy setup that cycles IP addresses across requests.

[Learn more](/content/glossary/rotating-proxy/index.html)

R

## SafetyBench Safety Benchmark

A benchmark suite for measuring large language model safety performance.

[Learn more](/content/glossary/safetybench-safety-benchmark/index.html)

S

## Safety Metric

Evaluation score for how safely an LLM refuses or handles harmful requests.

[Learn more](/content/glossary/safety-metric/index.html)

## Scikit-learn

Python toolkit for traditional ML algorithms and evaluation.

[Learn more](/content/glossary/scikit-learn/index.html)

S

## Script Injection Data Privacy Attack

Tests whether AI responses can include harmful scripts that enable XSS and data privacy breaches.

[Learn more](/content/glossary/script-injection-data-privacy-attack/index.html)

## Segmentation in Machine Learning

Grouping data into meaningful segments for modeling or CV.

[Learn more](/content/glossary/segmentation-in-machine-learning/index.html)

S

## Selective Sampling

Choosing informative samples for labeling to reduce annotation cost.

[Learn more](/content/glossary/selective-sampling/index.html)

S

## Self-Consistency Evaluation Metric

Measures agreement across multiple sampled answers or chain-of-thought paths.

[Learn more](/content/glossary/self-consistency-evaluation-metric/index.html)

S

## Semantic Router

Routes queries to tools or agents using semantic similarity or classifiers.

[Learn more](/content/glossary/semantic-router/index.html)

S

## Semi-supervised Learning

Learning from scarce labels combined with unlabeled examples.

[Learn more](/content/glossary/semi-supervised-learning/index.html)

S

## Sensitivity and Specificity in ML

TPR and TNR metrics describing binary detection performance.

[Learn more](/content/glossary/sensitivity-and-specificity-in-ml/index.html)

S

## Sentiment Analysis

NLP task that labels text as positive, negative, neutral, or by emotion.

[Learn more](/content/glossary/sentiment-analysis/index.html)

S

## Sentiment Analysis Metric

Measures accuracy or agreement of sentiment/polarity predictions.

[Learn more](/content/glossary/sentiment-analysis-metric/index.html)

## Seq2Seq Model

Encoder-decoder architectures transforming one sequence into another.

[Learn more](/content/glossary/seq2seq-model/index.html)

S

## Shadow Deployment

Serving a candidate model on production traffic for comparison without user impact.

[Learn more](/content/glossary/shadow-deployment/index.html)

S

## Shapley Values

Game-theoretic attributions explaining each feature's contribution.

[Learn more](/content/glossary/shapley-values/index.html)

S

## Shell Injection Excessive Agency Attack

Probes for shell injection when AI agents can execute system commands - a key excessive-agency risk.

[Learn more](/content/glossary/shell-injection-excessive-agency-attack/index.html)

## Six-Month Moratorium

2023 proposal to pause training of frontier models beyond GPT-4-class systems.

[Learn more](/content/glossary/six-month-moratorium/index.html)

S

## Sliding Window Attention

Attention pattern restricting each token to a fixed-size local window for efficiency.

[Learn more](/content/glossary/sliding-window-attention/index.html)

S

## Softmax Function

Normalizes logits into a probability distribution summing to one.

[Learn more](/content/glossary/softmax-function/index.html)

S

## SQL Injection Excessive Agency Attack

Tests whether AI agents with DB access can be prompted into unauthorized SQL.

[Learn more](/content/glossary/sql-injection-excessive-agency-attack/index.html)

## SSRF Excessive Agency Attack

Tests AI agents for prompt-induced SSRF through browsing or HTTP tools.

[Learn more](/content/glossary/ssrf-excessive-agency-attack/index.html)

## Stereotypes and Discrimination Harmful Content Attack

Tests whether models resist prompts that push stereotyped or discriminatory outputs.

[Learn more](/content/glossary/stereotypes-and-discrimination-harmful-content-attack/index.html)

## StrategyQA Reasoning Benchmark

Benchmark of yes/no questions needing implicit multi-step reasoning strategies.

[Learn more](/content/glossary/strategyqa-reasoning-benchmark/index.html)

S

## Summarization Metric

Evaluates summary factual alignment and coverage of essential source content.

[Learn more](/content/glossary/summarization-metric/index.html)

## Supervised Machine Learning

ML paradigm that learns a mapping from features to labels using labeled training data.

[Learn more](/content/glossary/supervised-machine-learning/index.html)

S

## Support Vector Machines (SVM)

Supervised models that separate classes with a maximum-margin hyperplane, often with kernels.

[Learn more](/content/glossary/support-vector-machines-svm/index.html)

S

## Surrogate Model

Cheap approximate model standing in for an expensive simulation or black-box function.

[Learn more](/content/glossary/surrogate-model/index.html)

S

## SVAMP Math Benchmark

Benchmark of varied arithmetic word problems for testing LLM math reasoning.

[Learn more](/content/glossary/svamp-math-benchmark/index.html)

S

## Sycophancy Hallucination Attack

Tests whether models flip facts to match user framing—sycophantic hallucination.

[Learn more](/content/glossary/sycophancy-hallucination-attack/index.html)

## Synthetic Data

Artificially generated data used for ML training, testing, and privacy-preserving validation.

[Learn more](/content/glossary/synthetic-data/index.html)

S

## Synthetic Data Generation

Process of fabricating datasets that mimic real data for ML and privacy use cases.

[Learn more](/content/glossary/synthetic-data-generation/index.html)

S

## Tabular Data

Data organized in rows and columns, as in spreadsheets and relational tables.

[Learn more](/content/glossary/tabular-data/index.html)

T

## TAP Prompt Injection Attack

Automated jailbreak via Tree of Attacks with Pruning (TAP) iterative prompt search.

[Learn more](/content/glossary/tap-prompt-injection-attack/index.html)

## Task Completion Metric

Measures whether an AI agent successfully completes its assigned end-to-end task.

[Learn more](/content/glossary/task-completion-metric/index.html)

## TensorFlow

Open-source deep learning library for training and serving neural networks.

[Learn more](/content/glossary/tensorflow/index.html)

T

## Test Set in Machine Learning

Data subset reserved to evaluate a final model on unseen examples.

[Learn more](/content/glossary/test-set-in-machine-learning/index.html)

T

## Text Statistics Metric

Measures fundamental text properties such as length, word count, and character count.

[Learn more](/content/glossary/text-statistics-metric/index.html)

## Tone Metric

Evaluates whether AI-generated text matches the intended tone and communication style.

[Learn more](/content/glossary/tone-metric/index.html)

## Tool Correctness Metric

Evaluates whether AI agents select and call the correct tools with appropriate arguments.

[Learn more](/content/glossary/tool-correctness-metric/index.html)

## Top-1 Error Rate

Fraction of cases where the top predicted class is incorrect.

[Learn more](/content/glossary/top-1-error-rate/index.html)

T

## Topic Classification Metric

Scores how correctly text is categorized into predefined topic labels.

[Learn more](/content/glossary/topic-classification-metric/index.html)

## Toxicity Metric

Scores whether LLM outputs contain offensive or harmful language for moderation and safety gates.

[Learn more](/content/glossary/toxicity-metric/index.html)

## Training Serving Skew

Difference between training-time and serving-time data that harms production performance.

[Learn more](/content/glossary/training-serving-skew/index.html)

T

## Transfer Learning

Applies pretrained knowledge to new related tasks, often by fine-tuning.

[Learn more](/content/glossary/transfer-learning/index.html)

T

## Transformer Models

Self-attention neural networks that underpin modern language and multimodal models.

[Learn more](/content/glossary/transformer-models/index.html)

T

## Transformer Neural Network

Attention-based neural network architecture for sequence modeling tasks.

[Learn more](/content/glossary/transformer-neural-network/index.html)

T

## Transliteration Prompt Injection Attack

Tests whether script/transliteration obfuscation bypasses AI content filters.

[Learn more](/content/glossary/transliteration-prompt-injection-attack/index.html)

## Tree-Based Models

Algorithms using decision trees for classification and regression tasks.

[Learn more](/content/glossary/tree-based-models/index.html)

T

## Tree of Thoughts

Prompting method that explores branching reasoning paths for harder multi-step tasks.

[Learn more](/content/glossary/tree-of-thoughts/index.html)

T

## TreeSHAP

Explains tree-based model predictions using Shapley value attributions.

[Learn more](/content/glossary/treeshap/index.html)

T

## Triplet Loss Function

Metric-learning loss comparing anchor, positive, and negative inputs.

[Learn more](/content/glossary/triplet-loss-function/index.html)

T

## True Positive Rate

Proportion of actual positives correctly identified by a classifier.

[Learn more](/content/glossary/true-positive-rate/index.html)

T

## Trulens

Open feedback-functions framework for evaluating and monitoring LLM applications.

[Learn more](/content/glossary/trulens/index.html)

T

## TruthfulQA Reasoning Benchmark

Assesses LLM ability to reason toward truthful answers amid common misconceptions.

[Learn more](/content/glossary/truthfulqa-reasoning-benchmark/index.html)

T

## TruthfulQA Safety Benchmark

Benchmark for LLM truthfulness against common online misconceptions and false beliefs.

[Learn more](/content/glossary/truthfulqa-safety-benchmark/index.html)

T

## t-SNE

Technique for projecting high-dimensional data into insightful 2D/3D visualizations.

[Learn more](/content/glossary/t-sne/index.html)

T

## Type 1 Error

Incorrectly rejecting a true null hypothesis; also known as a false positive.

[Learn more](/content/glossary/type-1-error/index.html)

T

## Type 2 Error

Failing to reject a false null hypothesis; also known as a false negative.

[Learn more](/content/glossary/type-2-error/index.html)

T

## Unauthorized Advice Misguidance Attack

Tests whether AI agents refuse out-of-scope regulated advice under pressure.

[Learn more](/content/glossary/unauthorized-advice-misguidance-attack/index.html)

## Uncertainty Quantification

Methods that estimate prediction reliability to support safer AI deployment decisions.

[Learn more](/content/glossary/uncertainty-quantification/index.html)

U

## Underfitting in Machine Learning

Model too simple to capture data patterns, weak on training and new data.

[Learn more](/content/glossary/underfitting-in-machine-learning/index.html)

U

## Unsupervised Learning

Models that infer structure from untagged data without predefined targets.

[Learn more](/content/glossary/unsupervised-learning/index.html)

U

## Validation Set in Machine learning

Data subset used to tune hyperparameters and select models before final testing.

[Learn more](/content/glossary/validation-set-in-machine-learning/index.html)

V

## Variational Autoencoder

Generative model introducing controlled randomness for data encoding and synthesis.

[Learn more](/content/glossary/variational-autoencoder/index.html)

V

## Vector Databases

Databases optimized for storing and querying high-dimensional embedding vectors.

[Learn more](/content/glossary/vector-databases/index.html)

V

## Vertical AI Agents

Industry-specialized AI agents designed for domain workflows and constraints.

[Learn more](/content/glossary/vertical-ai-agents/index.html)

V

## VGGNet

Deep convolutional network known for simplicity and depth in image recognition.

[Learn more](/content/glossary/vggnet/index.html)

V

## Vicuna Conversation Benchmark

Evaluates conversational AI on response quality, coherence, and helpfulness across turns.

[Learn more](/content/glossary/vicuna-conversation-benchmark/index.html)

V

## Vision Language Models

AI models combining computer vision and NLP for multimodal tasks.

[Learn more](/content/glossary/vision-language-models/index.html)

V

## Word Overlap Metric

Evaluates textual similarity via n-gram overlap methods such as BLEU and ROUGE.

[Learn more](/content/glossary/word-overlap-metric/index.html)

## XGBoost

Optimized gradient boosting library for speed and performance in ML tasks.

[Learn more](/content/glossary/xgboost/index.html)

X

## XSTest Harmful Content Attack

Evaluates AI on XSTest ambiguous questions for calibrated harmful-content refusals.

[Learn more](/content/glossary/xstest-harmful-content-attack/index.html)

## YOLO (Object Detection Algorithm)

Fast, accurate algorithm that detects objects in real time in a single pass.

[Learn more](/content/glossary/yolo-object-detection-algorithm/index.html)

Y

## Zero-Shot Learning

Models recognize or solve unseen categories using related prior knowledge.

[Learn more](/content/glossary/zero-shot-learning/index.html)

Z

[Previous](/content/glossary?db2722f8_page=6/index.html) [Next](/content/glossary?db2722f8_page=2/index.html)

No results

Get AI security insights in your inbox

##### No vulnerabilities found?
  We refund the assessment.

**Limited-Time Offer:** available until the end of September.

**Business email**\*

Company name\*

What specific AI use case or risk are you targeting?\*

Demo Video V2 from Giskard AI on Vimeo

Playing in picture-in-picture

Play

Settings

Speed

QualityAuto

2x

1.5x

1.25x

1x

0.75x

0.5x

Picture-in-Picture

Fullscreen

Hit ▶️ now for your full Hub demo! Full video in your email.