liveBook · Manning

Accuracy

Also called correctness

Accuracy is the degree to which a result agrees with a specified truth, reference, or criterion, commonly measured as the proportion of evaluated predictions that are correct.

How 49 Manning books explain it

Grokking Machine Learning

A classification metric equal to the proportion of evaluated data points for which the predicted label matches the correct label.

How it is used
Measures correct classifications to evaluate models and compare candidate models or decision-tree splits.
How deep it goes
Defines a correct-prediction ratio, works through counts, and examines misleading results; defers error-sensitive judgments to other metrics.
Examples
Uses patient diagnoses, spam emails, app recommendations, and a video-recommendation exercise.
What you can do after
Evaluate/understand it by calculating accuracy, comparing classifiers, and checking whether high accuracy conceals important errors.

Read how Grokking Machine Learning explains accuracy →

Vibe Engineering

Correctness is the degree to which software behavior satisfies its intended specification and real-world requirements.

How it is used
Uses accuracy to assess AI outputs against external standards and weigh reliability against engineering costs.
How deep it goes
Defines accuracy and works through context-size benchmarks, while distinguishing type-checking signals from proof of correctness.
Examples
Compares a medicine-risk classifier’s predictions and text-to-SQL overlap scores at different query-history lengths.
What you can do after
Evaluate and understand model accuracy using task-specific error measures and quality-versus-cost benchmarks.

Read how Vibe Engineering explains accuracy →

Human-in-the-Loop Machine Learning: Active learning and annotation for human-centered AI

A measure of an annotator's accuracy, estimated in this context by comparing label distributions across annotators.

How it is used
Uses accuracy to assess annotator reliability and decide which annotations to trust.
How deep it goes
Works through chance-adjusted ground-truth accuracy and a finite-annotator adjustment for predicted response distributions.
Examples
Calculates a 60% score for predicting that 65% of annotators choose a label when 40% do.
What you can do after
Evaluate annotator accuracy against ground truth or observed response distributions and select an annotation-confidence threshold.

Read how Human-in-the-Loop Machine Learning: Active learning and annotation for human-centered AI explains accuracy →

Effective Conversational AI

An accuracy rate is the proportion of evaluated predictions or answers that are correct.

How it is used
Measures how often evaluated chatbot predictions or answers are correct, providing a baseline for improvement.
How deep it goes
Defines the calculation, illustrates test selection and score interpretation, then directs diagnosis toward intent-level metrics rather than treating accuracy as a guarantee.
Examples
Uses a 79-of-100 classifier test and blind-test results after adding login and cancellation utterances.
What you can do after
Diagnose/debug with it by establishing a judged baseline, then inspecting poorly performing, high-volume intents and retesting changes.

Read how Effective Conversational AI explains accuracy →

Starting Data Analytics with Generative AI and Python

Classification accuracy is the proportion of observations whose predicted class matches the actual class.

How it is used
Evaluates review-score predictions to inform an e-commerce analysis.
How deep it goes
Demonstrates accuracy calculation on held-out predictions and pairs the result with a confusion matrix, without explaining the metric’s implementation.
Examples
Uses a customer review-score model that reports 43.03% accuracy on a test set.
What you can do after
Apply it via scikit-learn’s accuracy_score to compare predicted and actual test-set classes.

Read how Starting Data Analytics with Generative AI and Python explains accuracy →

Causal Inference for Data Science

Accuracy is the proportion of predictions that match the corresponding observed outcomes.

How it is used
Uses accuracy to judge predictive performance while separating prediction from causal efficacy.
How deep it goes
Defines accuracy and illustrates its causal limitation, then defers propensity-score evaluation to AUC.
Examples
Uses routine bread purchases, a perfectly accurate recommender, and treatment-assignment classifiers.
What you can do after
Evaluate predictive performance and distinguish accurate forecasts from evidence of causal impact.

Read how Causal Inference for Data Science explains accuracy →

Grokking Machine Learning, Second Edition

Accuracy is the proportion of evaluated examples for which a model produces the correct prediction.

How it is used
Evaluates classification models when comparing candidates and checking quality before deployment.
How deep it goes
Defines accuracy, works through numerical calculations, and contrasts its limits with precision and recall.
Examples
Uses a 875-of-1,000 test set, age-based decision stumps, disease testing, neural networks, and handwritten-digit recognition.
What you can do after
Evaluate/understand classification accuracy and recognize when error-specific metrics are needed.

Read how Grokking Machine Learning, Second Edition explains accuracy →

Statistics Every Programmer Needs

Model accuracy is the proportion of predictions that a model assigns correctly to the target outcomes.

How it is used
Evaluates how often binary classifiers match observed outcomes and supports test-set comparisons between decision trees and random forests.
How deep it goes
Defines a confusion-matrix formula, calculates a percentage manually, demonstrates scikit-learn scoring, and examines errors by class.
Examples
Calculates 86.78% accuracy for raisin classes and compares 61% decision-tree accuracy with 62% random-forest accuracy on fourth-down outcomes.
What you can do after
Apply it via scikit-learn `metrics.accuracy_score` to test labels and predictions, then inspect a confusion matrix for class disparities.

Read how Statistics Every Programmer Needs explains accuracy →

AI Governance

Accuracy is the data-protection principle that personal data be correct and, where necessary, kept up to date.

How it is used
Treats accuracy as a data-protection obligation and an operational risk in generated outputs.
How deep it goes
Explains architectural safeguards, escalation, measurement, and monitoring while noting that grounding and citations cannot guarantee correct conclusions.
Examples
Uses fabricated statements about real people, internal summaries, medical recommendations, and automated claim denials to illustrate accuracy risks.
What you can do after
Evaluate accuracy risks and apply controls for correction, escalation, measurement, and human review.

Read how AI Governance explains accuracy →

Natural Language Processing in Action, Second Edition

Accuracy is the proportion of evaluated predictions that are correct.

How it is used
Uses accuracy as an intuitive score for monitoring classifier training and evaluating predictions on held-out data.
How deep it goes
Defines accuracy through correct-prediction counts and implements binary calculations, but does not develop alternative evaluation metrics.
Examples
Measures toxicity classification, binary name-gender prediction, and RNN nationality prediction.
What you can do after
Evaluate model predictions by computing the correct fraction and comparing training with test or validation accuracy.

Read how Natural Language Processing in Action, Second Edition explains accuracy →

Classic Computer Science Problems in Swift: Essential techniques for practicing programmers

Accuracy is the proportion of evaluated examples that a classifier assigns to the correct category.

How it is used
Measures correct classifications when testing trained neural networks on iris and wine data.
How deep it goes
Implements the correct-count calculation in Swift and works through test results, without developing a broader metric comparison.
Examples
Uses iris classification at 9 of 10 correct and wine classification at 27 of 28 correct.
What you can do after
Apply it via Swift `validate()` to report correct counts and percentages, then compare results after tuning network settings.

Read how Classic Computer Science Problems in Swift: Essential techniques for practicing programmers explains accuracy →

Software Testing with Generative AI

Accuracy is a quality characteristic concerning whether software produces correct results and represents information correctly.

How it is used
Uses accuracy to clarify requirements and identify risks in software behavior and data handling.
How deep it goes
Develops accuracy through structured questions and code-risk analysis, stopping short of prescribing complete tests or fixes.
Examples
Demonstrates accuracy with timesheet requirements, half-hour durations, invalid entries, and Java duration-accumulation logic.
What you can do after
Evaluate generated questions and code risks to find unclear accuracy requirements and possible incorrect results.

Read how Software Testing with Generative AI explains accuracy →

Statistics Slam Dunk

Model accuracy is the proportion of instances that a classification model predicts correctly.

How it is used
Evaluates how often a binary classification model makes correct predictions, including NBA postseason qualification.
How deep it goes
Defines accuracy, gives two formulas, and calculates it from a confusion matrix; it stops short of treating accuracy as sufficient alone.
Examples
Uses an NBA postseason confusion matrix with 40 true positives, 19 true negatives, 11 false positives, and 10 false negatives.
What you can do after
Evaluate model correctness by deriving accuracy from predicted and actual binary outcomes.

Read how Statistics Slam Dunk explains accuracy →

Deep Learning with R, Third Edition

Accuracy is the proportion of evaluated predictions that exactly match their corresponding target labels.

How it is used
Monitors classification performance during training and evaluation and helps define project success.
How deep it goes
Defines accuracy, demonstrates Keras evaluation, and explains interpretation through validation curves and simple baselines.
Examples
Uses handwritten digits, 46-class Reuters classification, and a 90% majority-class baseline.
What you can do after
Apply it via Keras to monitor training and evaluate a classifier on test data.

Read how Deep Learning with R, Third Edition explains accuracy →

Data-Oriented Programming in Java

Correctness is the property that an implementation produces the behavior and results specified by its intended semantics and requirements.

How it is used
Defines behavior through laws and invariants so tests can verify domain correctness rather than implementation coverage.
How deep it goes
Develops equations, algebraic properties, domains, images, orderings, and integration checks, while avoiding a requirement for elaborate algebraic interfaces.
Examples
Uses duplicate-data cleanup, account-to-organization rules, task lifecycle transitions, and ledger-to-mirror synchronization tests.
What you can do after
Diagnose/debug with it by expressing domain laws and verifying implementations and dependencies against every relevant state.

Read how Data-Oriented Programming in Java explains accuracy →

Introduction to Generative AI, Second Edition

Accuracy is the proportion of predictions or outputs that are correct for a specified task.

How it is used
Highlights the risk of relying on plausible LLM responses in professional work.
How deep it goes
Warns that LLM outputs can be false without specifying how to measure accuracy.
Examples
Uses a legal brief containing invented lawsuits to illustrate inaccurate ChatGPT output.
What you can do after
Evaluate LLM responses by checking them against reliable sources, especially in specialized fields.

Read how Introduction to Generative AI, Second Edition explains accuracy →

Data-Driven Cybersecurity

Accuracy is the proportion of evaluated instances that a classification model labels correctly.

How it is used
Evaluates overall correctness when comparing cybersecurity classifiers and assessing threat-detection results.
How deep it goes
Defines the percentage of correct classifications and interprets test results, then qualifies them with class-level metrics rather than deriving the formula.
Examples
Compares random-forest and SVM login results and examines a neural-network report whose 0.63 accuracy masks zero recall for one class.
What you can do after
Evaluate/understand it by comparing test accuracy with class-level errors before judging a security classifier useful.

Read how Data-Driven Cybersecurity explains accuracy →

LLM Evaluation and Alignment, The Foundational Ideas

Accuracy is the proportion of evaluated outputs that match the correct answer or ground truth.

How it is used
Uses factual correctness to distinguish reliable language-model outputs from hallucinations.
How deep it goes
Distinguishes correctness from grounding and describes verification approaches, without specifying an accuracy calculation.
Examples
Uses financial-report summaries and document question answering to illustrate source-based verification.
What you can do after
Evaluate and understand whether claims are grounded and factually aligned, while recognizing that omitted answers require separate attention.

Read how LLM Evaluation and Alignment, The Foundational Ideas explains accuracy →

Architecting for Autonomy

Accuracy is the reliability of the premises, interpretations, propositions, and conclusions from which an agent reasons and acts.

How it is used
Treats accuracy as a control on information used in agent decisions and as a criterion for evaluating claims-triage performance.
How deep it goes
Defines accuracy beyond source values, identifies interpretation failures, and gives verification controls; the claims pilot examines consequences beyond an aggregate score.
Examples
Uses an illustrative claims-triage pilot and legal propositions tied to governing versions, delivery records, transcripts, and supporting authorities.
What you can do after
Evaluate accuracy risks by checking source and claim fidelity and assessing the downstream consequences of classification errors.

Read how Architecting for Autonomy explains accuracy →

Machine Learning for Tabular Data: XGBoost, Deep Learning, and AI

Accuracy is the proportion of classification predictions that match the corresponding true labels.

How it is used
Evaluates balanced binary or multiclass classifiers and compares model performance during tabular-data workflows.
How deep it goes
Defines correctness, demonstrates direct computation, and extends evaluation through cross-validation, but defers imbalanced-class alternatives.
Examples
Uses synthetic classification data, Scikit-learn cross-validation output, and Airbnb NYC price-category predictions.
What you can do after
Evaluate classification models with accuracy, cross-validation, and percentage-based prediction reports.

Read how Machine Learning for Tabular Data: XGBoost, Deep Learning, and AI explains accuracy →

Generative AI in Action

Accuracy is the proportion or degree of evaluated outputs that are correct according to a specified reference or criterion.

How it is used
Uses accuracy as one criterion for assessing whether generative AI outputs suit their intended applications.
How deep it goes
Outlines task-based evaluation, baseline comparison, and monitoring without specifying an accuracy calculation.
Examples
Compares a fine-tuned translation model with a baseline and examines content filters for false positives and false negatives.
What you can do after
Evaluate model performance using task-specific criteria, baseline comparisons, automated metrics, and human review.

Read how Generative AI in Action explains accuracy →

Outlier Detection in Python

Accuracy is a metric measuring the proportion of a model's predictions that match the correct outcomes.

How it is used
Guides detector and ensemble selection by comparing predictions with labeled, synthetic, or pseudo-ground-truth outcomes.
How deep it goes
Compares metrics and ensemble choices, but leaves metric selection dependent on the evaluation task.
Examples
Uses doped records to assess six detectors and weight their scores, and balances accuracy against correlation when selecting detectors.
What you can do after
Evaluate detector combinations against reference outcomes and use measured accuracy to select or weight ensemble members.

Read how Outlier Detection in Python explains accuracy →

LLM Customization and Fine-Tuning

The proportion of evaluated predictions or answers that are correct.

How it is used
Uses accuracy to compare prompting and adaptation choices on held-out tasks and to assess RAG answers.
How deep it goes
Explains test-set sizing, scoring methods, diminishing data gains, and comparisons with retrieval hit rate and faithfulness.
Examples
Illustrates accuracy with few-shot ticket routing, e-commerce extraction, and product-catalog tagging.
What you can do after
Evaluate and understand accuracy baselines, choose test-set sizes, and diagnose gaps between retrieval hits and correct RAG answers.

Read how LLM Customization and Fine-Tuning explains accuracy →

Grokking AI Algorithms, Second Edition

An accuracy score is the proportion of evaluated predictions that match their correct outcomes.

How it is used
Measures overall classification performance on unseen examples and helps guide model improvement.
How deep it goes
Defines overall correctness through a worked classification example, while noting other metrics reveal additional performance details.
Examples
Uses a ten-diamond cut-classification test set, with seven predictions matching their actual classes.
What you can do after
Evaluate/understand a classifier’s overall correctness by comparing its predictions with actual outcomes.

Read how Grokking AI Algorithms, Second Edition explains accuracy →

Unit Testing Principles, Practices, and Patterns

A metric of how reliably a test suite indicates whether the tested functionality is correct or broken, accounting for false positives and false negatives.

How it is used
Evaluates whether test results provide dependable guidance about code correctness.
How deep it goes
Defines four test outcomes and connects accuracy to false negatives, false positives, regression protection, and refactoring resistance.
Examples
Uses a four-quadrant test-result classification and a signal-versus-noise model.
What you can do after
Evaluate tests by their ability to find bugs without generating implementation-driven false alarms.

Read how Unit Testing Principles, Practices, and Patterns explains accuracy →

Inside Deep Learning: Math, Algorithms, Models

The fraction of predictions that match the true labels.

How it is used
Guides Optuna hyperparameter search as a validation score, then measures final performance on the test set.
How deep it goes
Defines accuracy as the fraction of correct predictions and uses it as a score, without detailing its computation.
Examples
Compares scratch, warm-start, and frozen transfer learning at barely over 50%, about 80%, and about 91% test accuracy.
What you can do after
Evaluate competing hyperparameters with validation accuracy and reserve test accuracy for final assessment.

Read how Inside Deep Learning: Math, Algorithms, Models explains accuracy →

Hidden Influences

The proportion of predictions that are correct among all predictions made by a system.

How it is used
Uses accuracy as a training objective for digit recognition and contrasts it with the goal of recommendation value.
How deep it goes
Compares perfect and random-guess performance and relates accuracy to error rate.
Examples
Uses digit recognition, random guesses among digits 0–9, and user ratings of movie recommendations.
What you can do after
Evaluate digit recognition against perfect and random-guess baselines and distinguish prediction success from recommendation value.

Read how Hidden Influences explains accuracy →

Financial AI in Practice

Accuracy is the proportion or degree of predictions that a model gets correct relative to the evaluated cases or reference outcomes.

How it is used
Evaluates fraud classification and market-direction predictions against their reference outcomes and practical baselines.
How deep it goes
Explains misleading accuracy under class imbalance and demonstrates conditional out-of-sample directional evaluation, while deferring trading friction and capital allocation adjustments.
Examples
Uses an all-legitimate fraud classifier, an always-up market baseline, and a 52.2% conditional directional result.
What you can do after
Evaluate and understand accuracy by checking the denominator, naive baseline, error types, and real-world trading limitations.

Read how Financial AI in Practice explains accuracy →

Build AI Drug Discovery Pipelines

Accuracy is the proportion of all classified instances that receive the correct predicted class.

How it is used
Evaluates classifier predictions in ligand-screening and CYP3A4 inhibition workflows.
How deep it goes
Uses training and validation results to illustrate accuracy, then warns that it is insufficient alone for imbalanced classes.
Examples
Uses hERG blocker predictions and CYP3A4 inhibition to compare model accuracy with validation results and dummy baselines.
What you can do after
Evaluate classifier scores and compare them with baselines and validation performance.

Read how Build AI Drug Discovery Pipelines explains accuracy →

Building Specialized AI Systems

Accuracy is the proportion of task outputs that are correct according to the relevant evaluation criteria.

How it is used
Guides systematic selection among AI approaches by setting the required performance level alongside consistency and knowledge needs.
How deep it goes
Frames accuracy as a solution-selection dimension but does not explain metric calculation or evaluation procedures.
Examples
Uses a four-quadrant decision framework and branches for training-data, retrieval, personalization, and domain-reasoning needs.
What you can do after
Evaluate application requirements and select an appropriately simple or specialized AI approach.

Read how Building Specialized AI Systems explains accuracy →

Inside AI

Predictive accuracy is the degree to which a model's predictions correspond to the outcomes that actually occur.

How it is used
Evaluates whether AI predictions are dependable enough for consequential decisions and business operations.
How deep it goes
Compares predictions with outcomes and discusses error consequences, but does not give a formal accuracy calculation.
Examples
Uses criminal-risk predictions and rejected online payments to show the human and commercial costs of errors.
What you can do after
Evaluate systems by considering prediction errors, bias, and the consequences of incorrect decisions.

Read how Inside AI explains accuracy →

How Large Language Models Work

Accuracy is the proportion of model outputs or predictions that are correct according to the evaluation criterion.

How it is used
Evaluates how often model predictions are correct.
How deep it goes
Defines accuracy through correct counts and explains why training uses cross-entropy instead.
Examples
Uses cancer predictions for 50 patients, including a case with 48 correct predictions, to illustrate discrete changes.
What you can do after
Evaluate/understand why accuracy measures success but cannot directly train an algorithm.

Read how How Large Language Models Work explains accuracy →

Build a Text-to-Image Generator (from Scratch)

Accuracy is the proportion of predictions that exactly match the corresponding ground-truth labels.

How it is used
Evaluates image classifiers by summarizing exact-label correctness over test predictions.
How deep it goes
Explains prediction extraction and batch aggregation with runnable PyTorch code, stopping short of alternative classification metrics.
Examples
Uses CIFAR-10 ViT evaluation at 0.5621 and balanced real-versus-fake ResNet50 evaluation at 0.9373.
What you can do after
Apply accuracy via PyTorch evaluation loops that compare argmax predictions with labels and report the aggregate result.

Read how Build a Text-to-Image Generator (from Scratch) explains accuracy →

Machine Learning System Design

Accuracy is the proportion of evaluated predictions that match the corresponding correct labels.

How it is used
Monitors batch predictions by comparing them with actual demand.
How deep it goes
Names prediction accuracy as a monitoring metric but does not define its calculation.
Examples
Contrasts an ugly AR-mask frame with inaccurate medical or transport results, and tracks batch predictions against actual demand.
What you can do after
Evaluate batch predictions by tracking accuracy against actual demand.

Read how Machine Learning System Design explains accuracy →

Data Without Labels

Model accuracy is the degree to which a predictive model produces correct outputs for the cases being evaluated.

How it is used
Checks data correctness before analysis and assesses clustering quality through compactness and separation.
How deep it goes
Defines data accuracy, recommends KPI checks, and describes clustering metrics including WCSS, intercluster distance, and silhouette values.
Examples
Uses annual sales completeness, a 2015–2019 study scope, business KPI comparisons, and a hypothetical million-customer clustering.
What you can do after
Evaluate clustering compactness and separation using distance-based measures and the silhouette value.

Read how Data Without Labels explains accuracy →

AI Agents and Applications

How it is used
Guides model selection and identifies a quality goal for improving retrieval-augmented generation.
How deep it goes
Compares accuracy qualitatively with speed and cost, but provides no measurement procedure.
Examples
Contrasts customer-support chatbots with legal analysis tools and points to RAG workflow stages.
What you can do after
Evaluate model choices by balancing accuracy against latency and cost for application requirements.

Read how AI Agents and Applications explains accuracy →

Learn Generative AI with PyTorch

The fraction or percentage of predictions that match the true labels in a classification task.

How it is used
Evaluates classification predictions against true labels on held-out test data.
How deep it goes
Calculates accuracy by thresholding binary probabilities or selecting the highest-scoring class, then averaging label matches.
Examples
Uses Fashion MNIST images to classify ankle boots versus T-shirts and to predict ten clothing categories.
What you can do after
Evaluate a trained classifier by comparing thresholded binary or argmax multicategory predictions with test labels.

Read how Learn Generative AI with PyTorch explains accuracy →

Build a Large Language Model (From Scratch)

The fraction or percentage of predictions that match the correct class labels in a classification task.

How it is used
Measures spam-classification performance before, during, and after supervised fine-tuning.
How deep it goes
Explains final-token argmax, batch aggregation, device handling, and evaluation limits, but uses cross-entropy for optimization.
Examples
Uses spam-versus-non-spam labels, two-class logits, batch loaders, and reported training, validation, and test percentages.
What you can do after
Apply it via a PyTorch accuracy loader to compare model predictions with target labels across datasets.

Read how Build a Large Language Model (From Scratch) explains accuracy →

Reinforcement Learning from Human Feedback

Training accuracy is the proportion of training examples for which a model produces the correct output.

How it is used
Distinguishes overfitting from proxy-objective failure by contrasting training and held-out accuracy.
How deep it goes
Defines training accuracy as the share of training examples answered correctly and contrasts it with held-out accuracy.
Examples
Uses training-versus-held-out accuracy to distinguish memorization from reward-model over-optimization.
What you can do after
Evaluate whether degrading held-out accuracy reflects overfitting rather than a proxy objective diverging from user satisfaction.

Read how Reinforcement Learning from Human Feedback explains accuracy →

Fight Fraud with Machine Learning

Accuracy is the proportion of predictions that match the actual class labels.

How it is used
Evaluates classification systems, especially fraud and synthetic-voice detectors, after predictions are compared with labeled test data.
How deep it goes
Defines and calculates the proportion of correct predictions, implements test-loop evaluation, and stops short of treating it as sufficient for imbalanced fraud data.
Examples
Uses fraud emails, fake-voice clips, and illicit Bitcoin transactions to demonstrate accuracy calculations and practical limitations.
What you can do after
Evaluate classification models via PyTorch and judge when accuracy requires support from imbalance-aware metrics.

Read how Fight Fraud with Machine Learning explains accuracy →

"Looks Good to Me"

The proportion of AI-generated findings or suggestions that are correct according to the relevant evaluation criteria.

How it is used
Guides decisions about when AI-generated review findings or code changes are safe to trust.
How deep it goes
Measures accuracy numerically and operationally through confidence levels, thresholds, observation, and human overrides.
Examples
Uses AI flags and misses, rated outputs, pull-request decisions, and a 31% comment-to-code translation result.
What you can do after
Evaluate AI review performance and choose adoption thresholds while preserving human control over incorrect suggestions.

Read how "Looks Good to Me" explains accuracy →

Effective Data Analysis

Data accuracy is the degree to which data correctly represents the real-world process or phenomenon it claims to describe.

How it is used
Uses accuracy as a component of data quality to support reliable information for decisions and, in some organizations, legal compliance.
How deep it goes
Explains real-world representation, implausible values, duplicates, and proactive checks across data-pipeline steps.
Examples
Uses prohibited negative values and duplicate records as accuracy problems, alongside checks on forecast dates and weather values.
What you can do after
Apply pipeline checks to identify implausible values and duplicate data before they cause downstream problems.

Read how Effective Data Analysis explains accuracy →

Retrieval Augmented Generation, The Foundational Ideas

Accuracy is the proportion of evaluated system outputs judged correct according to a specified criterion.

How it is used
Evaluates whether systems return correct answers or retrieved content under a specified benchmark criterion.
How deep it goes
Distinguishes answer correctness from Recall@k and MRR, then reports strict and semantic retrieval accuracy.
Examples
Compares baseline RAG at 54.12% Hit Rate@5 with A-RAG at 62.35% strict and 69.41% semantically adjusted accuracy.
What you can do after
Evaluate retrieval systems by comparing accuracy with baselines, semantic scoring, and query latency.

Read how Retrieval Augmented Generation, The Foundational Ideas explains accuracy →

Machines that Think

Accuracy is the degree to which outputs correctly correspond to the intended facts, labels, or answers.

How it is used
Measures answer correctness in model evaluation and reveals how prompt changes can affect classification performance.
How deep it goes
Explains reward asymmetry and prompt sensitivity, but does not provide a formula for calculating accuracy.
Examples
Uses a dolphin mislabeled as a fish and studies that vary prompt formatting, punctuation, and example order.
What you can do after
Evaluate model behavior by checking answers against known labels and comparing results across prompt variations.

Read how Machines that Think explains accuracy →

AI Model Evaluation

A measure of how often a model's outputs match desired or correct outcomes.

How it is used
Validates whether an LLM judge’s verdicts agree with trusted, known-correct labels before production.
How deep it goes
Distinguishes agreement with labels from run-to-run stability, without specifying a formula or calculation procedure.
Examples
Contrasts a judge that is often correct but noisy with one that is consistent yet wrong.
What you can do after
Evaluate whether an LLM judge matches known-correct labels; do not treat that result as evidence of rerun stability.

Read how AI Model Evaluation explains accuracy →

The Art of AI Product Development

Accuracy is the degree to which a model's outputs match the correct or intended results.

How it is used
Uses accuracy as a criterion for evaluating language models and aligning their outputs with specific user needs.
How deep it goes
Recommends benchmarks, tailored metrics, and real-world validation but does not define a calculation method.
Examples
Uses factual accuracy in a content-generation product and task accuracy for specialized use cases.
What you can do after
Evaluate model outputs with standardized benchmarks, task-specific metrics, and real-world feedback.

Read how The Art of AI Product Development explains accuracy →

AI Agents in Action, Second Edition

Routing accuracy is the proportion of inputs assigned to the appropriate agent or processing route.

How it is used
Evaluates structured outputs against known targets when correctness has a clear right-or-wrong answer.
How deep it goes
Explains when accuracy applies and contrasts it with rubric-based evaluation, without detailing metric calculations.
Examples
Uses classification labels, extracted fields, optimization decisions, summaries, explanations, plans, and creative content as output examples.
What you can do after
Evaluate structured outputs with standard metrics and select other methods when correctness is not clearly defined.

Read how AI Agents in Action, Second Edition explains accuracy →

Transfer Learning for Natural Language Processing

Accuracy is the proportion of evaluated samples whose predicted labels match their actual labels.

How it is used
Evaluates and compares classification performance on held-out samples.
How deep it goes
Defines the correct-prediction ratio and discusses metric choice, but does not detail accuracy’s limitations beyond class balance.
Examples
Uses email, IMDB sentiment, and fake-news classification results to illustrate accuracy.
What you can do after
Evaluate classifier predictions by comparing predicted labels with actual labels.

Read how Transfer Learning for Natural Language Processing explains accuracy →

Build an AI Agent (From Scratch)

Accuracy is the proportion of evaluated answers or predictions that match the correct reference outcomes.

How it is used
Measures how often models answer evaluation problems correctly.
How deep it goes
Defines accuracy as correct answers out of 20 and compares model scores, without elaborating further on metric calculation.
Examples
Uses 20 GAIA level-1 problems, a 5/20 tool-free ceiling, and model score comparisons.
What you can do after
Evaluate model performance by comparing correct-answer counts across a fixed set of problems.

Read how Build an AI Agent (From Scratch) explains accuracy →

Also covered in

What to know first

Related concepts

Questions about accuracy

What is accuracy?

Accuracy is the degree to which a result agrees with a specified truth, reference, or criterion, commonly measured as the proportion of evaluated predictions that are correct.

Which Manning books explain accuracy?

Grokking Machine Learning, Vibe Engineering, Human-in-the-Loop Machine Learning: Active learning and annotation for human-centered AI, Effective Conversational AI, Starting Data Analytics with Generative AI and Python, Causal Inference for Data Science, Grokking Machine Learning, Second Edition, Statistics Every Programmer Needs and 41 more.

What should I know before learning accuracy?

Manning books that teach accuracy usually introduce language model, classifier, prompt and fine-tuning first.

What is the difference between accuracy and fidelity?

Accuracy is the degree to which a result agrees with a specified truth, reference, or criterion, commonly measured as the proportion of evaluated predictions that are correct. Fidelity is the degree to which an output, model, or representation accurately reflects or is supported by its relevant source, target, evidence, or behavior.

What kinds of accuracy are there?

Kinds of accuracy include incident classification accuracy, per-word accuracy and transcription accuracy.

What is accuracy also called?

Accuracy is also called correctness.

These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.